Systems and methods for prediction of performance parameters for multicomponent layered devices

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating one or more performance parameters for a multicomponent layered device using device configuration and material inputs. One of the methods includes receiving configuration and material inputs for the device; extracting features from the inputs to generate configuration features and materials features; and processing the features using a machine learning (ML) model to generate one or more performance parameters for the device. Another method includes determining an importance metric for each configuration and material feature with regard to performance parameters generated by the ML model to enable more informed user design decisions of devices. A further method includes adjusting the trainable parameters of the ML model based on user-specified data, without modifying the ML model's configuration, allowing a user to tailor the ML model more accurately to the user's needs and device types.

Description

Attorney Docket No. 17367-0160W01SYSTEMS AND METHODS FOR PREDICTION OF PERFORMANCE PARAMETERS FOR MULTICOMPONENT LAYERED DEVICESCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of priority to U.S. Application No. 63 / 750,529, filed on January 28, 2025, the contents of which are hereby incorporated by reference.TECHNICAL FIELD

[0002] The specification relates to integrating data from various sources, including computational models, experimental studies, literature reports, and intrinsic properties such as stoichiometry and physical descriptors (collectively features or descriptors), with machine learning algorithms to predict the operational performance of multicomponent layered devices.BACKGROUND

[0003] Techniques exist that attempt to use machine learning (ML) methods to quantify the relationship between device performance and one or more of its material or architectural components. Such techniques emphasize components of certain ‘critical’ layers where the principal physicochemical mechanisms of action behind device operation occur, such as the emission layer for organic light-emitting diodes (OLEDs) and the absorber layer for perovskiteabsorber based solar cells (PSCs). These techniques generally entail predicting physicochemical properties of materials behind device operation, (e g., the HOMO, LUMO, and / or singlet-triplet gap of emitter molecules in the emission layer of an OLED), which relate to a device performance parameter (e.g., the external quantum efficiency of an OLED). But these techniques may suffer in that they (1) preclude other layers which may significantly alter device operation and (2) do not naturally lend to deciding actionable changes to altering devices. Further, these techniques may be unable to “featurize” (i.e., convert into structured, machine-readable features for processing by a machine learning model) layer and device arrangements, components, properties, etc. in a complete and arbitrary way. This inability to featurize any and all types of configurations or materials of a device may cause issues such as: (1) inability to generalize across multiple device types; (2) limited scalability to large and diverse candidate device sets; (3) limited characterization of all relevant layers or relevant material components of these in layers; (4) limited featurization of chemical structure; (5) over reliance on property data from eitherAttorney Docket No. 17367-0160W01slow and costly experiments or computational simulations; and (6) inability to handle datasets that are either incomplete or small (but complex).SUMMARY

[0004] This specification describes, among other aspects, technology implemented as computer programs on one or more computers in one or more locations that process configuration and material inputs of a multicomponent layered device to predict values of one or more performance parameters for the multicomponent layered device.

[0005] The technology described herein includes a comprehensive framework which allows treatment of all layers of all kinds of devices without sacrificing user flexibility or and speed of prediction. For each layer, as well as the device as a whole, this specification enables encoding of features of any type of properties, whether they be structural, physical, chemical, mechanical, industrial, and so on.

[0006] A multicomponent layered device (or alternatively, a multilayered device), refers to a device composed of a series of layers of chemical material, with each layer material selected to meet certain property criteria critical to its role in the overall operation of the multilayer device. The operational output, stability, and efficiency of a multicomponent layered device depends on the structure, composition, and amount of each chemical layer, as well as the manner and order the layers are arranged. The operational output, stability, and efficiency of a device may be quantified by one or more performance parameters measured during its operation.

[0007] The design of novel multicomponent layered devices involves exploring novel compositions and arrangements of their constituent chemical layers to optimize these performance parameters to some set of desired values. Exploration of novel device configurations requires significant time, labor, money, and / or materials, due to the costs of the experimental fabrication and analyses involved.

[0008] Intelligent selection of novel material components, as well as layer sizes, film morphologies, and sequential ordering in the device is also challenging due to the complexity of the physicochemical processes occurring in the device during operation. Understanding how the modification of certain components of the device configuration affects different aspects of operational performance is nontrivial. The more complex the device, the more complex this assessment becomes. Due to the high complexity of multicomponent layered devices, the designAttorney Docket No. 17367-0160W01of multicomponent layered devices is a significantly complicated, time-consuming, and costly endeavor.

[0009] High-throughput virtual screening (HTVS) strategies are commonly used to efficiently explore the properties and performance of individual materials or subsystems for use in multicomponent layered devices across diverse configurations and material compositions. By leveraging low-cost computational models built from diverse datasets, HTVS predicts the properties of individual materials or subsystems, but it does not provide direct predictions of the device performance as a whole.

[0010] The data underlying HTVS can originate from diverse sources, including experimental methods and computational techniques such as quantum mechanics (QM), molecular dynamics (MD), and / or other simulation methods. While experimental data can provide empirical validation, it is often limited in scope with respect to chemistry, due to the time and effort involved to perform new experiments. Simulation methods, on the other hand, offer broader coverage and higher consistency, enabling predictions of chemical and material properties across a wide spectrum of device components. Moreover, they allow the establishment of quantitative relationships between chemical structure and physicochemical properties.

[0011] While these approaches excel at screening and optimizing individual component materials, no effective HTVS strategies currently exist for addressing the complexities of full device architectures. Currently, physics-based methodologies are fundamentally ill-suited for incorporating the complete structure of multilayer devices into their predictive computational frameworks. Direct application of QM or MD to all atoms or molecules of a general multilayer device architecture is computationally prohibitive, limiting their practical use. Simulation methods have been developed to model device phenomena, e.g., those that utilize drift diffusion methods, which model charge transport and recombination. However, these methods operate under various simplifying assumptions which make general application challenging, including, but not limited to issues stemming from the allowed charge distributions, allowed modes of interfacial charge transfer, allowed optical pathways, and the need to employ classical treatments of charge mobility and recombination. Moreover, they are computationally intensive. In sum, it is not practical to run simulations on a substantially large number of devices, of varying complexity, in an efficient manner.Attorney Docket No. 17367-0160W01

[0012] Effective HTVS strategies are more feasible for the evaluation of individual material components in a device, as material property values can be more feasibly obtained through experiments or computational predictions. However, inferring the performance parameters of a complete multilayered device from these isolated material properties is highly challenging, particularly with regard to a quantitative relationship. Further, sophisticated analyses are needed to include interactions between the interfaces of device components which are absent from direct isolated material calculations. Furthermore, for chemically complex materials, modeling efforts may still be relegated to truncated systems and / or isolated, gas-phase molecules, which may fail to capture the bulk-phase behavior or other environmental interactions crucial for device operation. In sum, relying solely on material property values is insufficient for comprehensive device screening and evaluation. This is particularly true for architectural modifications beyond so-called critical layers that contain materials directly involved in the primary physicochemical process driving the device.

[0013] In either case, the general application of computational methods, e.g., QM and / or MD, to facilitate HTVS requires significant expertise in the field. While various simplifications to the computational treatment and / or the chemical system may make these approaches more tenable for a general user, they may compromise the generalization of workflows and / or the veracity of the predictions. As a result, the reliance on sophisticated QM and / or MD simulations can pose a significant barrier to implementing effective HTVS strategies for many multicomponent layered devices. This highlights the need for a novel method, as proposed in this specification, which can leverage these techniques, when necessary, but does not depend on them as a prerequisite.

[0014] Formulation machine learning (formulation ML) enables prediction of properties or behaviors of mixtures of molecules given specific composition ratios. These chemical species can be represented in a variety of ways, e.g., as graphs or molecular property descriptors.However, formulation ML is typically constrained to single, mixed systems, and lacks a capability to account for the structural and organizational complexity of multicomponent layered devices. Specifically, it cannot fully accommodate the variability in the arrangement and interactions of layers, the number and composition ratios of distinct materials within each layer, or the specific orientation and ordering of these layers, all of which critically influence device performance.Attorney Docket No. 17367-0160W01

[0015] This specification describes machine learning (ML) technologies that can be trained using representations of a multicomponent layered device to enable reliable predictions of performance parameters. That is, this specification describes techniques to process configuration and material inputs for a multicomponent layered device to generate values for one or more performance parameters for the multicomponent layered device. The computational cost to generate performance parameters is generally low, the performance parameters are accurate, and no user expertise is necessary.

[0016] As used in this specification, the term “featurization” with respect to the described techniques corresponds to operations that extract certain features from certain input, the features being formatted for processing by a machine learning model. For example, featurization includes extracting configuration features from the configuration input and extracting material features from the material input, with both sets of features formatted for processing by a machine learning model.

[0017] As one advantage, the described techniques can encode the important structural and chemical property features of all components and interfaces of a multicomponent layered device to enable predictions of operational performance parameters for the entire device, not just properties of individual layers or materials. That is, by processing both configuration input, which consists of both comprehensive device configuration parameters (i.e., parameters associated with operation or construction of the device as a whole, e.g., the operational voltage of the device) and individual layer configuration parameters (i .e., parameters associated with the layer specification, e.g., for each layer, a layer thickness or volume value;, composition ratio values; layer fabrication methods; nanostructure size, morphology and porosity; device area and form factors; and so on) and materials input (i.e., parameters associated with the layer composition, e g., for each layer, structural properties, electronic properties, ensemble properties, morphological properties, and so on), the described techniques can use representations of a multicomponent layered devices to enable reliable predictions of performance parameters of these devices.

[0018] As another advantage, the described techniques leverage the results of QM and / or MD simulations for predicting performance parameters efficiently. That is, the described techniques can efficiently balance QM and / or MD simulations by performing them on-the-flyAttorney Docket No. 17367-0160W01when compute time permits, leveraging stored results when available, and / or using pre-trained surrogate models to approximate values as needed.

[0019] As a consequence of the above two advantages, the described techniques enable a user to perform rapid, accurate performance prediction of multicomponent layered devices. Additionally, a user has flexibility to employ featurizations that are as detailed as is necessary or achievable for the desired accuracy. In addition to a single system performing the described techniques to be able to provide rapid, accurate performance prediction of multicomponent layered devices, multiple systems performing the described techniques in parallel could further speed up performance prediction of multiple multicomponent devices by each executing in parallel.

[0020] As another advantage, the described techniques can enable users to perform accurate data-driven multicomponent layered device design by highlighting which materials and properties correlate with device performance parameters.

[0021] For example, device configuration inputs, device material inputs, and respective performance parameters can be combined with feature importance techniques to determine which materials and properties correlate with device parameter values, enabling accurate data-driven multicomponent layered device design.

[0022] As another example, a user varying the configuration and material inputs and observing resultant performance parameters allows the user to explore new devices beyond the scope of devices that have been already manufactured. As a result, the user can optimize the structure and composition of layered devices through the use of the described techniques without being limited to manufacturing the device.

[0023] As another advantage, the described techniques can enable users to train models using their own data while handling the complexity of ML model training and configuration (e.g., neural network architecture design, model hyper-parameter choices, and so on) for the user, offering flexibility and personalization compared to fixed pre-trained models, without requiring the user to have specific expertise. Further, the techniques can enable a user to maintain control over their proprietary data while using a ML model training procedure or ML model configuration that is shared among many.

[0024] By enabling rapid predictions of user-defined devices, the described techniques pave the way for improved future developments of multicomponent layered devices, e.g.,Attorney Docket No. 17367-0160W01OLEDs, transistors, photovoltaics, capacitors, batteries, and so on. Therefore, the described techniques address existing limitations with ML techniques with regard to simulating full devices, while maintaining low compute cost and a low barrier-to-entry, while still enabling accurate predictions of performance parameter(s).

[0025] The details of one or more embodiments of the subject matter of this specification are set forth in the accompanying drawings and the description below.

[0026] According to a first aspect there is provided a method of using a machine learning (ML) model to predict performance parameters of a device comprising two or more layers that includes receiving, at a computer, configuration input for the device, wherein the configuration input comprises device-wide parameters and / or layer parameters for each of the two or more layers; receiving, at the computer, material input for the device, wherein the material input comprises composition values of each of the two or more layers; extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model, the configuration features comprising values derived from the configuration input formatted for processing by the ML model; extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model, the material features comprising values derived from the material input formatted for processing by the ML model; and processing the configuration features and the material features using the ML model to generate one or more performance parameters, the ML model having been trained using a training dataset comprising a plurality of training examples, each training example comprising values for one or more target performance parameters and respective configuration features and material features, wherein the ML model is trained to process the respective configuration features and respective material features and output a value for each of the one or more performance parameters.

[0027] In some implementations, training the ML model includes obtaining the training dataset comprising training examples; for each training example, processing the respective configuration features and materials features and generating one or more performance parameters; evaluating an objective function using the one or more performance parameters for each training example; and updating the ML model trainable parameters to optimize the objective function.Attorney Docket No. 17367-0160W01

[0028] In some cases, the composition values of each of the two or more layers comprises one or more chemical structures or one or more chemical properties.

[0029] In some cases, one or more of the chemical properties are experimental measurements.

[0030] In some cases, one or more of the chemical properties are computed electronic properties.

[0031] In some cases, one or more of the chemical properties are computed ensemble and / or morphological properties.

[0032] In some cases, one or more of the chemical properties are computed using pretrained surrogate models.

[0033] In some cases, one or more chemical properties are as described in any one of the above cases or implementations.

[0034] In some cases, one or more chemical properties explicitly encode the physical and chemical interactions and / or relationships between layers.

[0035] In some cases, the configuration input further comprises the overall device-wide structural and operational parameters, as well as the layer-specific parameters of dopant composition ratios and layer fabrication methods for each of the two or more layers.

[0036] In some cases, the method further includes maintaining data of material input values, and wherein extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model comprises: determining whether the material input is missing one or more values; retrieving the one or more missing values from maintained data; and computing the one or more missing values, wherein the computing is performed for any missing values not retrieved from the maintained data.

[0037] In some cases, computing the missing values of the material input comprises using a pre-trained surrogate model.

[0038] In some cases, extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model includes determining whether the configuration input is missing one or more values; and in response to determining that the configuration input is missing one or more values, imputing configuration input missing values.Attorney Docket No. 17367-0160W01

[0039] In some cases, imputing the configuration input missing values comprises using a pre-trained imputer neural network.

[0040] In some cases, the method further includes determining, for each feature of the configuration features and material features, all possible subsets of configuration features and material features that do not include the feature; processing, for each feature and for each subset, the feature and the subset of features using the ML model to generate one or more performance parameters; processing, for each feature and for each subset, the subset of features using the ML model to generate one or more performance parameters; determining, for each feature and for each subset, a marginal contribution for each performance parameter as the difference between performance parameter generated from processing the feature and the subset of features using the ML model and the respective performance parameter generated from processing the subset of features using the ML model; determining, for each feature and for each marginal contribution, a weight based on the size of the subset of features used to determine the marginal contribution; and determining an importance metric, which quantifies the significance of each feature with regard to predictions of each performance parameter, as the sum of respective marginal contributions, each scaled by its respective weight.

[0041] In some cases, the method further includes: adjusting the trainable parameters of the ML model based on user-specified data, without modifying the ML model’s configuration; and generating one or more user-specified performance parameters using the ML model.

[0042] In some cases, the two or more layers are layers of a device selected from the group consisting of a light emitting diode, transistor, a photovoltaic cell, a capacitor, a battery, a sensor, a flexible display, and a multilayer printed circuit board.

[0043] In some cases, the ML model is a model selected from the group consisting of an artificial neural network, a tree-based model, a support vector machine, a k-nearest neighbor model, and a generalized linear model.

[0044] According to a second aspect there is provided the methods of the first aspect performed by a system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one more computers to perform operations of the respective method.

[0045] According to a third aspect, there is provided the methods of the first aspect performed by one or more non-transitory computer storage media storing instructions that whenAttorney Docket No. 17367-0160W01executed by one or more computers cause the one more computers to perform operations of the respective method.

[0046] These embodiments demonstrate the efficacy of the specification by providing reliable predictions for multiple important performance properties as metrics in device design and fabrication. The example device space covers multiple types with different ‘horizontal’ functionalities: OLEDs and PSCs, as well as ‘vertical’ architectures: single-unit and tandem OLEDs.

[0047] Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.Attorney Docket No. 17367-0160W01BRIEF DESCRIPTION OF THE DRAWINGS

[0048] FIG. 1 shows an example device performance forecasting system.

[0049] FIG. 2 is a flow diagram of an example process for generating performance parameters for a device.

[0050] FIG. 3 is a flow diagram of an example process for training a machine learning model to generate one or more performance parameters.

[0051] FIG. 4 shows an example of the performance of the described techniques for single-unit organic light-emitting diodes.

[0052] FIG. 5 shows an example of the performance of the described techniques for tandem organic light-emitting diodes.

[0053] FIG. 6 shows an example of the performance of the described techniques for single-unit perovskite-absorber based solar cells.

[0054] FIG. 7 shows an example of importance metric values for certain configuration and material features of single-unit organic light-emitting diodes that are identified to be significant with regard to predicting the device performance parameter: external quantum efficiency.

[0055] FIG. 8 shows an example of importance metric values for certain configuration and material features of single-unit perovskite-absorber based solar cells that are identified to be significant with regard to predicting the device performance parameter: power conversion efficiency.

[0056] FIG. 9 shows an example of the performance of the described techniques for single-unit organic light-emitting diodes, for tandem organic light-emitting diodes, and for single-unit perovskite-absorber based solar cells.

[0057] FIG. 10 shows an example of user-specified data interfacing with a device performance forecasting system.

[0058] FIG. 11 shows an example computer system.

[0059] Like reference numbers and designations in the various drawings indicate like elements.Attorney Docket No. 17367-0160W01DETAILED DESCRIPTION

[0060] FIG. 1 shows an example device performance forecasting system 100. The device performance forecasting system 100 is an example of a system implemented as computer programs on one or more computers in one or more locations, in which the systems, components, and techniques described below can be implemented.

[0061] The system 100 uses a machine learning (ML) model 116 to predict performance of a device that includes two or more layers (i.e., a multicomponent layered device). That is, the system 100 receives input for the device (i.e., a configuration input 102A and a material input 102B) and predicts values for performance parameter(s) 104 of the device (i.e., metrics for performance, efficiency, and output of operating the multicomponent layered device).As some specific examples, the system 100 can be used to predict characteristic performance parameters of organic light emitting diodes (OLEDs) and perovskite-based absorber solar cells (PSCs). These are devices composed of a series of layers in unique arrangements that are composed of different chemical compositions to perform opposing functions.

[0062] OLEDs are devices composed of two or more layers of amorphous films of organic compounds and organometallic complexes, which emit light when an electric field is applied driving current through the device via terminal electrodes. The system 100 can predict performance properties of OLEDs, such as external quantum efficiency, power efficiency, current efficiency, color indices, and electroluminescence band- structure, using the ML model 116 and inputs for the device.

[0063] PSCs are devices composed of combinations of two or more layers of amorphous organic films and / or polycrystalline inorganic compounds of various nanostructures (particles, rods, bundles) adjacent to both sides of a perovskite crystalline layer, which absorbs light to stimulate electrical current. The system 100 can predict performance properties of PSCs, such as the power conversion efficiency, short-circuit current, open-circuit current voltage, and fill factor, using the ML model 116.

[0064] Generally, the multicomponent layered device can be any of a variety of devices. That is, the multicomponent layered device can be any device with two or more stacked materials or components, where each layer serves a specific function. In addition to vertically stacked structures, the multicomponent layered device can also include lateral structures. Such devices may include materials or components that are arranged in the horizontal plane, whereAttorney Docket No. 17367-0160W01functional regions are spatially separated along the horizontal plane in addition to the vertical regions. Some examples of multicomponent layered devices include light emitting diodes (e.g., organic light emitting diodes), transistors, photovoltaics (e.g., solar panels utilizing perovskite crystalline fdms, dye solutions, and / or organic fdms as absorbers), capacitors, batteries, sensors, flexible displays, multilayer printed circuit boards, and so on.

[0065] For certain device types, multiple device units may be vertically stacked, providing multiple inputs / outputs to enhance certain performance properties. Device architectures with one unit are often referred to as ‘single-unit’ or ‘single-junction’ devices, whereas those with multiple vertical stacked units are referred to as ‘tandem’ or ‘multi -junction’ devices. For example, ‘tandem OLEDs’ consist of two or more OLED devices, each with their respective emission layers and layer structure, that may be connected via unique layers such as charge generation layers. Tandem OLEDs incorporate multiple output light sources that allow for increased efficiency and brightness.

[0066] As a specific example of application of the prescribed techniques to multi-unit devices, the system 100 can be used to predict the performance parameters of tandem OLEDs. These properties include those cited above for single-unit OLEDs.

[0067] During operation, the system 100 receives, at a computer, a configuration input 102A for the device, wherein the configuration input includes layer-specific parameters for each of the two or more layers and device-wide parameters for the whole system.

[0068] For example, referring to the above to the single-unit and tandem OLED device examples, the system 100 can receive a configuration input 102A for the devices that involve descriptors for individual layers, such as their thickness, as well as descriptors for the device as a whole, such as the operational voltage for the entire OLED device (i.e., voltage level at which the device is designed to operate at, e.g., voltage needed to drive current through the device).

[0069] As an additional example, referring to the above PSC device example, the system 100 can receive a configuration input 102A for devices that involve descriptors for individual layers, such as the nanostructures of crystalline or films (e.g., particle size, shape, and porosity), as well features for the device as a whole, such as the device area (i.e., surface area of the device which absorbs light). Generally, the configuration input 102A includes any of a variety of descriptors or values relevant to the configuration specification of the individual layers of a device or aspects of the device as a whole. The two examples cited above are not exclusive toAttorney Docket No. 17367-0160W01their respective device types; such descriptors or values relevant to the configuration specification of the individual layers may be employed for any device to which the descriptors or values may be relevant.

[0070] For example, in addition to the thickness value for each of the two or more layers, the configuration input 102A can include other values relevant to the configuration specification of the device layers, such as composition ratios or dopant weight percentages (i.e., proportion of a specific dopant material in a given layer), which is relevant to (1) the emissive layer of OLEDs because composition values determine the efficiency, brightness, and color of emitted light from that layer and (2) the electron transport layer of PSCs because composition values determine the charge mobility and energy level alignment of that layer. As another example, the values relevant to the configuration specification of the device layers can include film fabrication method values for layers (e.g., vacuum deposition or solution processing), which influence each layer’s morphology and microstructure (e.g., how common defects are in a layer, how crystalline the layer is, and so on).

[0071] The system 100 also receives, at the computer, a material input 102B for the device, wherein the material input 102B includes composition values of each of the two or more layers.

[0072] Generally, the material input 102B includes any of a variety of descriptors or values relevant to the composition (i.e., chemical make-up) of the device layers.

[0073] For example, the material input 102B can include, for each layer, one or more chemical structures or one or more chemical properties. The material input may encode several classes of material components and mixtures existing in different media environments. Particular examples of materials include small organic, inorganic, and organometallic molecules; polymers; and nanostructures such as nanoparticles, nanorods, nanowires, and nanosheets; and so on.Particular examples of media environments include bulk solids, amorphous films, polycrystalline films, colloids, pure liquids, gels, solutions, suspensions, membranes, and so on.

[0074] As particular examples, the chemical structure can include Simplified Molecular Input Line Entry System strings (i.e., SMILES strings, i.e., ASCII strings representing the structure of chemical molecules), Morgan fingerprints (i.e., a standardized cheminformatics representation of the structure of a molecule), three dimensional geometric specification of atoms of a molecular or repeating periodic structure (e.g., xyz-coordinates, internal coordinates, latticeAttorney Docket No. 17367-0160W01parameters, and so on), nanostructure descriptors (e.g., size, morphology, and porosity), and / or percent composition of chemical mixtures, isomerism, stoichiometric ratios, and so on.

[0075] As particular examples, the chemical properties can include chemical hardness, molecular orbital levels, ground- and excited-state energy levels, band structure values, H0M0-LUMO gap, fluorescence, phosphorescence, reorganization energy, relaxation time, decoherence time, optical gaps,, spin multiplicity, conductivity, heat of formation, bonddissociation energies, strain, electronegativity, acidity, oxidation states, reactivity, electrochemical potential, solubility, polarity, electron density, vapor pressure, melting point, glass-transition temperature, surface tension, thermal conductivity, permeability, and so on. The configuration input 102A can be composed of chemical properties at many sizes, ranging from atomic scale to film scale.

[0076] In some cases, the chemical structure, or chemical properties are quantities that are experimentally derived (i.e., are experimental measurements). In other cases, these quantities are previously derived through theory, e.g., through QM or MD simulations (i.e., computational implementations of chemical and physical theories to determine structural or chemical properties, e.g., computed electronic properties or ensemble properties).

[0077] For example, the chemical structure, or chemical properties can include the values for the radial distribution function for molecules in amorphous films. This radial distribution function may be determined from experiments (e.g., using neutron or x-ray scattering data) or through theory (e.g., as computed by molecular dynamics simulations).

[0078] The system 100 extracts, by the computer and from the configuration input 102A, features from the configuration input 102A to generate configuration features 112 for inputting into the ML model 116, the configuration features 112 including values derived from the configuration input 102A formatted for processing by the ML model 116.

[0079] Simultaneously, the system 100 extracts, by the computer and from the material input 102B, features from the material input 102B to generate material features 114 for inputting into the ML model 116, the material features 114 including values derived from the material input 102B formatted for processing by the ML model 116.

[0080] That is, the system 100, processes the raw, unstructured data, i.e., values that make up the configuration input 102A and material input 102B, into structured formats that theAttorney Docket No. 17367-0160W01ML model 116 can process effectively, i.e., the configuration features 112 and materials features 114 respectively.

[0081] For example, for an input value that is a text string, e.g., a SMILES string, the system 100 can tokenize the string, e.g., using word embeddings, map the token sequence to embedding vectors that captures chemical semantic meaning, e.g., using a string encoder neural network, and pool the embedding vectors to a final fixed sized vector.

[0082] As another example, for an input value that is a text string, e.g., a SMILES string, the system 100 can convert the string into a 2D structure, e.g., with a cheminformatics library, e.g., RDKit, then featurize the structure, e.g., using Morgan fingerprint or any other molecular structure-based encoding technique, to embed the structure as a bit-vector.

[0083] As another example, for an input value that is categorical, e.g., structural input values such as (1) ‘layer type’, e.g., ‘substrate-layer’, ‘anode layer’, ‘emissive-layer’, and so on; or (2) ‘nanostructure’ type, e.g., ‘nanoparticle’, ‘nanowire’, ‘nanosheet’ and so on The system 100 can one-hot encode the categorical values (i.e., convert to binary vectors).

[0084] As another example, for an input value that is a real continuous or integer number, e.g., band gap value, electron affinity value, oxidation states, spin-multiplicity, and so on, system 100 can use the values directly as features. However, in some cases, the system 100 transforms these values through any of a variety of transformations (e.g., a linear transformation to scale the values, or a clip function that restricts the value to specified range).

[0085] As described above, in this specification, the term “featurization” corresponds to operations in which the device performance forecasting system 100 extracts, from the configuration input, features to generate configuration features, and extracts, from the material input, features to generate material features, with both sets of features formatted for processing by the ML model 116 (e.g., tokenization / embedding of strings, molecular fingerprints, one-hot encoding of categorical values, normalization or other transformations).

[0086] In some cases, the configuration input 102A and / or material input 102B are missing values, indicating, e.g., a user’s intent, for the system 100 to determine these values. The system 100 can determine the missing values, e.g., through imputing the missing values (e.g., using an imputer 106, e.g., a neural network imputer, a nearest neighbor imputer, a multivariate imputer, and so on), executing calculations (e.g., executing on-the-fly QM or MD simulations),Attorney Docket No. 17367-0160W01leveraging pre-trained models to predict the values (e.g., pre-trained surrogate models 110), or retrieving the values from databases (e.g., maintained data 108).

[0087] Further details of the system 100 determining missing values are described below.

[0088] In some cases, the system 100 further processes an initial extracted set of features using an encoder neural network to create a final set of features that serve as the set of features (i.e., configuration features 112 and material features 114).

[0089] The system 100, processes the configuration features 112 and the material features 114 using the ML model 116 to generate one or more performance parameters 104.

[0090] The ML model 116 can include any of a variety of types of ML models. Examples of possible types of ML models include, but are not limited to, artificial neural networks, treebased models (e.g., decision tree, random forests, gradient boosting models), support vector machines, k-nearest neighbors models, and generalized linear models (e.g., linear regression, logistic regression, and so on).

[0091] As a particular example, the ML model 116 can be an artificial neural network in which the artificial neural network can have any of a variety of neural network architectures. That is, the ML model 116 can be an artificial neural network and can have any appropriate architecture in any appropriate configuration that processes configuration features 112 and materials features 114 to generate performance parameters 104, including fully connected layers, convolution layers, recurrent layers, attention-based layers, and so on, as is appropriate.

[0092] Generally, the performance parameter(s) 104 can be any of a variety of characteristics of a device, e.g., performance and spectral characteristics of a device. The target performance parameters can depend upon the type of multicomponent layered device of interest. For example, the performance parameters for an OLED device can be (i) external quantum efficiency, (ii) internal quantum efficiency (iii) power efficiency, (iv) current efficiency, (v) color indices (e.g., CIE(x), and CIE(y)), (vi) electroluminescence band-structure (e.g., EL(max), EL(FWHM)), (vii) operational lifetimes (e.g., T50, T70, T90), (viii) turn-on voltage, and so on. As another example, the performance parameters for a PSC device can be (i) power conversion efficiency, (ii) short-circuit current, (iii) open-circuit current voltage, (iv) fill factor, and so on.

[0093] Prior to using the ML model 116 to generate performance parameter(s) 104, the system 100 or another training system trains the ML model 116 on a training dataset.Attorney Docket No. 17367-0160W01

[0094] In some cases, the system 100 first pre-trains the ML model 116 on a training dataset and then fine-tunes the ML model 116 on a different training dataset, e g., a training dataset provided by a user.

[0095] Further details of training the ML model 116 are described below.

[0096] The system 100 can use the performance param eter(s) 104 to select and control device fabrication. That is, the system 100 can use the performance parameters 104 of devices predicted by ML model 116 to directly influence fabrication processes of those devices.

[0097] As two particular examples, ML model 116 can be used in the process of (1) designing emitter molecules and optimizing their doping concentrations in host materials, as employed within some desired architecture of an OLED device and (2) choosing appropriate nanostructures for the electron transport layers of a PSC device, as employed in alongside some desired perovskite absorber layer formulation.

[0098] For example (1), device fabricators can explore emitters in a high-throughput screening procedure involving the enumeration of multiple candidate devices where the identity and concentration of the emitter differs, while the other layer materials and configurations are held constant. The emitter molecule and weight percent may be described via a SMILES string and decimal value, respectively. The SMILES strings of emitters in devices with optimally predicted performance directly point to molecular structures that are worth exploring experimentally for device assessment. With direct knowledge of the molecular structure, design fabricators can focus on emitters that are synthesizable via cost-effective routes. In addition, the importance metric values obtained from model training highlight significant chemical functionalities that should be considered when constructing a candidate emitter set.Subsequently, once a viable emitter has been identified, alterations to other layers (e.g., materials, size, fabrication process, and so on) to search for other optimizations. The example OLED performance evaluation described below in this specification utilizes a training set with many different emitter molecules, some with adjusted weight percentages.

[0099] For example (2), device fabricators may explore how different nanostructures of one or more electron transport layers influence the performance of solar cells constructed with a specific desired perovskite formulation in the absorber layer. As particular examinations, fabricators can explore (i) the efficiency benefits employing mesoporous polycrystalline film, with varying thicknesses, as scaffold layers for perovskite absorber layer, and / or (ii) theAttorney Docket No. 17367-0160W01nanostructure of material in the film. For examination (i), fabricators can enumerate devices where the electron transport layer either precludes a mesoporous layer, includes a mesoporous layer, or employs a mesoporous layer stacked on top of one or more compact layers; comparing the predicted performance of these device configurations drives fabrication of the electron transport layer. The thicknesses of these layers may also be adjusted to see what optimal size the mesoporous layer should be. Simple categorical descriptors may be employed for novel mesoporous layers to assign a general value, or more explicit 3D structural information (e.g., volume, density, porosity, periodicity, obtained from microscopy measurements or periodic DFT calculations) may be provided for more well-studied layers. Using similar exploratory means, fabricators may also pursue examination (ii) to investigate the benefits from employing different constituent nanostructures beyond standard nanoparticle and grain structures. Featurization may describe the film as being composed of nanorods, nanobundles, nanorods, or other non-regular structures. This featurization may also include additional descriptors for the 3D structure to explore how to optimize the layer preparation. The example PSC performance evaluation described below in this specification utilizes a training set with multiple different electron transport layer configurations.[000100] So, in some implementations, the system 100 determines that the generated one or more performance parameters 104 satisfy one or more design criteria. The system 100 then fabricates the device using one or more device fabricators, and the system 100 controls the one or more device fabricators based on the configuration input 102A and the material input 102B.[000101] Some examples of design criteria include target performance metrics specific to the device architecture being optimized. Additionally, design criteria may include fabrication constraints, such as a maximum material cost, a toxicity threshold for constituent materials, or a “synthesizability” score indicating the likelihood that a generated molecular structure can be successfully manufactured.[000102] For example, in the context of OLEDs, design criteria may include a minimum External Quantum Efficiency (EQE) threshold (e.g., >20%), specific CIE color coordinates, or a maximum driving voltage. In the context of PSCs, design criteria may include a minimum Power Conversion Efficiency (PCE).[000103] To facilitate control of one or more device fabricators based on the configuration input 102A and the material input 102B, the system 100 can include a fabrication interface thatAttorney Docket No. 17367-0160W01translates the configuration input 102A and material input 102B into machine-readable instructions (e.g., deposition protocols, temperature profiles, or flow rate settings) for the device fabricators. The “design criteria” may function as a threshold filter; for instance, the system 100 may be configured to trigger fabrication only if the predicted power conversion efficiency (PCE) exceeds a predetermined percentage or if a predicted device lifetime exceeds a specific duration. Once a candidate device configuration satisfies these criteria via the ML model 116 prediction, the system 100 retrieves the specific parameters defined in the inputs (e.g., layer thickness, doping concentration, or chemical composition) and maps them to specific control signals for the fabrication hardware.[000104] The system’s 100 ability to fabricate devices based on configuration input 102A, material input 102B and performance parameters 104 allows for an autonomous, closed-loop discovery process of novel devices. By iteratively generating candidate configuration inputs and material inputs, predicting their resulting performance, and physically fabricating only those devices correspond to performance parameters that the ML model 116 identifies as high-value targets, the system 100 drastically reduces material waste and experimental time.[000105] FIG. 2 is a flow diagram of an example process 200 for generating performance parameters for a device. For convenience, the process 200 will be described as being performed by a system of one or more computers located in one or more locations. For example, a device performance forecasting system, e.g., the device performance forecasting system 100 of FIG. 1, appropriately programmed in accordance with this specification, can perform the process 200.[000106] The system receives, at a computer, a configuration input for the device, wherein the configuration input includes device-wide parameters and layer parameters for each of the two or more layers (step 202).[000107] As described above, the configuration input can further include any of a variety of descriptors and or values relevant to the configuration specification of the device layers.[000108] In some cases, the configuration input includes overall device-wide structural and operational parameters, as well as layer-specific parameters for one or more layers.[000109] For examples of each, device-wide operational parameters may include the turnon voltage of an OLED device (i.e., minimum voltage at which emission starts), the area of PSC device (i.e., the surface area of the light-absorbing surface where sunlight is converted into electricity), or the form factor of a battery (i.e., overall shape of the device).Attorney Docket No. 17367-0160W01[000110] Layer specific parameters may include material layer descriptions such as the size and composition ratios as well as nanostructure specification, and layer fabrication methods. For example, for layer fabrication methods, one may encode values corresponding to different methodologies, including vapor based methods (e.g., vacuum thermal evaporation, chemical vapor deposition, atomic, layer deposition), solution processing based methods (e.g., spin coating, inkjet printing), patterning based methods (e.g., transfer printing). Similar encoding approaches provide description of different nanostructures materials may adopt within a film, including nanoparticle, nanowire, nanosheet, nanorod, or nanobundle. Accompanying descriptors may also be provided to account for varying sizes and arrangements of these structures (e.g., the volume, density, porosity, periodicity, percentage free volume ignoring the nanostructures, and so on). The system receives, at the computer, a material input for the device, wherein the material input includes composition values of each of the two or more layers (step 204).[000111] As described above, the materials input can include any of a variety of descriptors and or values relevant to the composition (i.e., chemical make-up) of the device layers, e.g., one or more chemical structures or one or more chemical properties.[000112] In some cases, the material composition values of each of the two or more layers includes one or more chemical structures or one or more chemical properties.[000113] For example, chemical structure inputs (e.g., SMILES strings, xyz-coordinates, internal coordinates, lattice parameters, nanostructure descriptors, composition ratios, and so on) may be featurized and be employed as standalone descriptors, or in combination with various physicochemical properties, such as those described below. As particular examples, a material in a given layer with increasing complexity using (1) Morgan fingerprint; (2) Morgan fingerprint with either electronic property descriptors (e.g., HOMO, LUMO, S1T1 gap and so on) or ensemble / morphological property descriptors (e.g., radial distribution function values, glasstransition temperature, and so on), or interlayer interaction property descriptors (e.g., H0M0-LUMO gap, relative dipole moment orientation, and so on); or (3) Morgan fingerprints inclusive of all of electronic property descriptors, ensemble / morphological property descriptors, and interlayer interaction property descriptors.[000114] In some cases, the one or more of the chemical properties are experimental measurements.Attorney Docket No. 17367-0160W01[000115] For example, oxidation potential, reduction potential, electronic band gap, optical band gap, delayed exciton lifetime, delayed fluorescence lifetime, prompt fluorescence lifetime, phosphorescence lifetime, wavelength and bandshape for maximum absorbance and emission, extinction coefficient, intersystem crossing rate, reverse intersystem crossing rate, photoluminescence quantum yield, refractive index, Abbe number, complex permittivity, dielectric constant, nanoparticle grain size, surface contact angle, vapor pressure, melting point, glass-transition temperature, decomposition temperature, specific heat, heat of vaporization, charge mobility, electronic conductivity, thermal conductivity, permeability, solubility, viscosity, diffusivity, and so on.[000116] In some cases, the one or more of the chemical properties are computed electronic properties.[000117] For example, these electronic properties include frontier molecular orbital energy levels (e g., HOMO, HOMO-1, HOMO-2, LUMO, LUMO+1, LUMO+2, and so on), static electric dipole moment, polarizability, molecular and film transition dipole moment, atomic charges, electrostatic potential values, hole reorganization energy, electron reorganization energy, triplet reorganization energy, gap between the energies of low-lying singlet states (Sn, 0 < n < 2) and the ground triplet-state energy, heat of formation, bond-dissociation energies, electron! c / vibronic excitation energies, specific heat, heat of vaporization. Note that above-cited experimental electronic properties that may also be computed by simulation, including oxidation potential, reduction potential, electronic band gap, optical band gap, delayed exciton lifetime, delayed fluorescence lifetime, prompt fluorescence lifetime, phosphorescence lifetime, wavelength and bandshape for maximum absorbance and emission, extinction coefficient, intersystem crossing rate, reverse intersystem crossing rate, photoluminescence quantum yield, refractive index, Abbe number, complex permittivity, and dielectric constant. Also note that any of these values may be supplied to ML model 116 via pre-trained surrogate models.[000118] In some cases, the one or more of the chemical properties are computed ensemble and / or morphological properties.[000119] For example, these ensemble and / or morphological properties include packing density, percentage free volume, Hansen solubility parameter, radius of gyration of the molecule, average cluster number and size (for representative dimers, trimers, and so on), bin of values for the distribution of the radial distribution function of clusters, bin of values for theAttorney Docket No. 17367-0160W01distribution of order parameter for the moment of inertia and dipole moment (relative to system), bin of values for the distribution of values for force field energy terms from derived energies (e.g., angle terms, dihedral terms, Coulomb terms, bond terms, van Der Waals terms, and so on). Note that above-cited experimental ensemble properties may also be computed by simulation methods, including vapor pressure, melting point, glass-transition temperature, decomposition temperature, specific heat, heat of vaporization, charge mobility, electronic conductivity, thermal conductivity, permeability, solubility, viscosity, diffusivity. Also note that any of these values may be supplied to ML model 116 via pre-trained surrogate models.[000120] In some cases, the.one or more of the chemical properties are computed using pretrained surrogate models. Further details of the surrogate models are described below.[000121] In some cases, the one or more chemical properties explicitly encode the physical and chemical interactions and / or relationships between layers.[000122] For example, these may include structural descriptors, such as surface contact angles, the counts and geometrical parameters (e.g., distance, angles) of different types of interlayer non-covalent interactions (e.g., hydrogen bonds, pi-pi stacking, and so on), changes in the uniformity of a surface layer before and after interaction with the adjacent surface layer. These may also include electronic descriptors, such as the gap between the HOMO and LUMO of adjacent layers as well as the relative orientation of the transition dipole moment order parameters of the adjacent layers. Any such values may be derived from experiment, simulation, or some combination of both.[000123] In some cases, the one or more chemical properties includes more than one of the above described chemical properties.[000124] The system can receive the inputs (i.e., configuration input and materials input) from a user or another system through any of a variety of methods, e.g., through a network connection, such as a cloud-based network, the internet, or a local network.[000125] For example, the system can receive the inputs from an end-to-end device design system. That is, the system can be integrated into a greater workflow for designing devices and can receive inputs and provide performance parameters to other entities belonging to the workflow.Attorney Docket No. 17367-0160W01[000126] As another example, the system can present an interface to a user, e.g., by establishing a network connection with an end-user device, prompt the user for inputs, and receive through the interface and therefore the network, the inputs.[000127] As a particular example, the user can provide to the system as inputs the number and order of layers of a multilayer device, descriptors of the configuration of each layer (e.g., thickness or volume, fabrication method, composition ratio, nanostructure morphology and porosity, etc.), and what the components of each of the layers are.[000128] The system extracts, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model, the configuration features including values derived from the configuration input formatted for processing by the ML model (step 206). That is, as described above, the system processes the raw, unstructured data, i.e., values of various types, e.g., string, category, numeric, etc., that make up the configuration input into structured formats that the ML model can process effectively.[000129] In some implementations, the system determines whether the configuration input is missing one or more values and, in response to determining that the configuration input is missing one or more values, imputes the configuration input missing values. In other words, the system can impute missing values of the configuration input, e.g., determining the missing value by mean imputation or median imputation or determining the missing value to be a predetermined value.[000130] In some cases, the system can impute the configuration input missing values using an imputer, e.g., a pre-trained imputer neural network.[000131] As a specific example, the system can utilize a neural network trained to predict and fill in missing values, i.e., imputation. The training process begins with mean imputation, where the missing values in the dataset are initialized to the mean of the observed (non-missing) values. During training, the neural network leverages both the observed non-missing values and the current set of imputed values as inputs to predict updated values for the missing entries. The process is iterative: the neural network refines the imputed values by repeatedly replacing them with its updated predictions. As the training progresses, the predicted values converge toward more accurate imputed values that align with the underlying data patterns. This iterative refinement continues until one or more stopping criteria are satisfied, such as reaching aAttorney Docket No. 17367-0160W01predefined maximum number of iterations or observing that the changes between consecutive predictions fall below a specified threshold. Once the training process is complete, the trained neural network imputer can be used to predict and impute missing values in new datasets, leveraging its learned understanding of the relationships between the non-missing and missing data.[000132] As a specific embodiment, the neural network imputer architecture ican be configured as a fully-connected multi-layer perceptron (MLP) or dense neural network, that includes an input layer, an output layer, and a collection of hidden layers. In one implementation, the architecture includes three hidden layers, each including 100 neurons and utilizing a nonlinear activation function, such as a Rectified Linear Unit (ReLU). The system may further employ a data augmentation module configured to generate multiple augmented data copies wherein missing values are randomly introduced into the feature set.[000133] The training of the neural network imputer can be performed through a sequential, nested-loop process. In each training epoch, the system can execute an internal iterative loop to refine the imputed values for the training batch before calculating the loss and updating the network weights. Specifically, within a single epoch, the network generates predictions, updates the imputed entries, and re-injects these updated values as inputs for the next internal iteration. This internal refinement can continue until a specific convergence condition is satisfied.[000134] To ensure robustness and numerical reliability of the imputation process, the system can incorporate stability and convergence enhancement measures. In one or more embodiments, the iterative refinement logic can be configured to apply a damping factor to each update cycle to prevent erratic oscillations in the predicted values. Specifically, rather than replacing a missing entry entirely with the network’s raw prediction, the system can calculate a weighted update, for example, by determining an average of the current predicted value and the existing imputed value from the previous iteration. Furthermore, to prevent the propagation of out-of-distribution values or numerical divergence, the system may implement a feature-specific constraint mechanism. In this configuration, the resulting imputed value is subjected to a clipping operation by the system, where the value is constrained to a predefined range.Specifically, a range defined by the minimum and maximum observed values for the corresponding feature column. This ensures all imputed data remains physically consistent with the observed data.Attorney Docket No. 17367-0160W01[000135] During the iterative refinement of this specific embodiment, the system can guide the optimization of the model using a composite loss function that includes a Mean Squared Error (MSE) loss and an iteration penalty. This penalty can be configured to apply a numerical penalty to the number of internal iterations required to reach a stable state within a single training epoch.[000136] The system can update the network weights by performing a backpropagation operation to compute gradients of the composite loss with respect to the model parameters. These weights are then adjusted using an Adaptive Moment Estimation (Adam) optimizer.[000137] The system extracts, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model, the configuration features including values derived from the material input formatted for processing by the ML model (step 208). That is, as described above, the system processes the raw, unstructured data, i.e., values of various types, e.g., string, category, numeric, etc., that make up the material input into structured formats that the ML model can process effectively.[000138] In some implementations, the system maintains data of material input values to use for extracting features from the material input. Particularly, when there are missing values in the materials input, the system will determine the missing values to be the corresponding values maintained in the data.[000139] In other words, the system can maintain data of material input values. Then, when the system extracts from the material input, features from the material input to generate material features for inputting into the ML model, the system can determine whether the material input is missing one or more values. The system can then retrieve the one or more missing values from maintained data. Afterwards, the system can compute the one or more missing values, where the system computes missing values for those not retrieved from the maintained data.[000140] In cases that the missing values are not in the maintained data, the system can compute what the missing values are.[000141] For example, consider a materials input that is missing the value of the triplet excitation energy for a polymer component of an emissive layer of an OLED device. The system can determine that this value is missing in the input and, in response, check the maintained data for the missing value. If the maintained data includes the missing value of the triplet excitationAttorney Docket No. 17367-0160W01energy for the polymer component of an emissive layer of an OLED device, the system determines the missing value to be that found in the maintained data.[0001421 Now consider the same example materials input missing value but, for this example, the maintained data did not include the missing value. The system, in response to determining the material input value is not in the maintained data, can compute the missing value using, e.g., using QM or MD simulations. For example, the system can initiate QM simulations to determine the ground state and lowest triplet state electronic energies for the polymer and determine the triplet excitation energy as the difference of those energies.[000143] In some cases, the system determines the details of the QM and / or MD simulations used to compute the missing values. In other cases, a user determines the details of those calculations. For example, a user can be prompted by the system through an end-user device to configure the details of a workflow of QM and / or MD simulations.[000144] In some implementations, to compute the missing value(s) of the material input, the system can use an existing, pre-trained ML model, e.g., a pre-trained neural network, to generate these values. That is, one or more ML models, previously trained on similar chemical systems, can be used to predict one or more materials input values. These models serve as advantageous ‘surrogate’ providers of QM and / or MD simulation results, thus enabling them to be used as input features, without incurring the time or compute cost of explicitly performing QM and / or MD simulations.[000145] The surrogate ML models can include any of a variety of types of ML models, e.g., artificial neural networks, tree-based models (e.g., decision tree, random forests, gradient boosting models), support vector machines, k-nearest neighbors models, or generalized linear models.[000146] In some cases, after determining the values of the missing values of the material input, the system stores the determined values in the maintained data for future use.[000147] The maintained data can be maintained through any of a variety of methods, e.g., cloud-based databases, on-premises databases, networked databases, on-machine databases, etc.[000148] Whether using maintained data, on-the-fly calculations, or surrogate neural networks to determine material input missing values, the system can leverage the non-missing input values to determine the missing values.Attorney Docket No. 17367-0160W01[000149] For example, if the chemical structure Morgan fingerprint of a layer is not missing, but the respective triplet excitation energy for the layer is missing, the system can (i) use the Morgan fingerprint as a query to lookup the respective triplet excitation energy in maintained data, (ii) process the Morgan fingerprint to determine workflow setups for on-the-fly QM and / or MD simulations to determine the triplet excitation energy, or (iii) process the Morgan fingerprint using a surrogate neural network to generate the triplet excitation energy.[000150] The above described steps 202-208 show that the configuration and material inputs are straightforward for non-expert users to define a multicomponent layered device to make predictions of performance parameters. Because the system imputes and determines any missing values of inputs, expert and non-expert users alike can define device architectures by providing as many inputs as they are comfortable doing. Consequently, expert and non-expert users alike can generate many different device architectures to conduct high-throughput screening and data post-processing.[000151] The system processes the configuration features and the material features using the ML model to generate one or more performance parameters (step 210).[000152] As described above, prior to using the ML model, the system trains the ML model.[000153] Further details of training the ML model are described below.[000154] FIG. 3 is a flow diagram of an example process 300 for training an ML model to generate one or more performance parameters. For convenience, the process 300 will be described as being performed by a system of one or more computers located in one or more locations. For example, a device performance forecasting system, e.g., the device performance forecasting system 100 of FIG. 1, appropriately programmed in accordance with this specification, can perform the process 300.[000155] The system or another training system trains the ML model by repeatedly updating the learnable parameters of the ML model using a training dataset of the system. That is, the system can repeatedly perform the following described example process using training examples that each include respective features, i.e., configuration features and materials features associated with a multicomponent layered device, and respective target performance parameter(s), i.e., the performance parameters associated with the multicomponent layeredAttorney Docket No. 17367-0160W01device, to train an ML model from scratch, i.e., train from randomly initialized parameters, or fine-tune, i.e., further train.[0001561 The system obtains the training dataset which includes training examples (step 302). The system can receive the training dataset from any of a variety of appropriate sources, e.g., a user, another system, system data repository, and so on. Generally, the collection of multicomponent layered devices included in the training dataset can have varying complexity (e.g., a device can have two layers with one component each or multiple layers with multiple components each).[000157] Generally, the system determines a plurality of training examples from the training dataset that each include respective features, i.e., configuration features and materials features associated with a multicomponent layered device, and respective target performance parameter(s), i.e., the performance parameters associated with the multicomponent layered device. Generally, the system determines the features for each training example by preprocessing raw, unstructured data into structured formats as described above. That is, the system pre-processes the training dataset to featurize the raw data. The pre-processing can include one-hot encoding categorical values, embedding string values, normalizing continuous numerical values, and so on.[000158] In some implementations, prior to pre-processing the training dataset, the system uses an imputer to impute configuration input missing values and trains a surrogate model to predict QM and / or MD simulations to determine material input missing values. Additionally, training the surrogate model can further include initializing system maintained data of material input values with the QM and / or MD simulations used to train the surrogate model.[000159] The system, for each training example, processes the respective configuration features and materials features and generates one or more performance parameters (step 304).[000160] That is, the system generates performance parameter(s) for each training example by processing at least the input features (i.e., configuration features and materials features) associated with the training example using the ML model.[000161] The system evaluates an objective function using the one or more performance parameters for each training example (step 306).[000162] The objective function can include a loss for each training example and any additional regularization terms to improve ML model generalization.Attorney Docket No. 17367-0160W01[000163] One example loss function to compute the loss for a training example is the squared error loss function associated with the training example target performance parameter(s) and respective generated performance parameter(s). More specifically, for this example, the squared error loss is the square of the difference of the generated performance parameter(s) with the respective target performance parameter(s). For this example, an example objective function can be the mean squared error, i.e., the mean of the squared error loss for all training examples.[000164] The system updates the ML model trainable parameters to optimize the objective function (step 308).[000165] The system can update the ML model trainable parameters to optimize the objective in any variety of ways, e.g., gradient based method, evolutionary algorithm -based method, Bayesian optimization, grid search, etc.[000166] For example, the system can optimize the objective function by minimizing the loss of one or more training examples described above using any of a variety of gradient descent techniques (e.g., batch gradient descent, stochastic gradient descent, or mini -batch gradient descent) that include the use of a backpropagation technique to estimate the gradient of the loss with respect to ML model learnable parameters and to update the learnable parameters accordingly.[000167] Generally, the system repeats the above steps (304-308) until one or more criteria are satisfied (e.g., the system performs a pre-determined number of iterations, the updates to the learning parameters no longer exceed a pre-determined magnitude of change, a metric regarding a validation dataset exceeds a pre-determined value, and so on).[000168] FIGS. 4-6 show examples 400-600 of the performance of the described techniques. That is, in these examples 400-600, the described techniques train and execute machine learning (ML) models to predict performance parameter(s) of a respective device that includes two or more layers (e.g., single-unit OLED devices, tandem OLED devices, and perovskite solar cells (PSCs)).[000169] FIG. 7 and FIG. 8 show examples 700 and 800 of configuration and material features with the largest importance metric values with regard to predicting device performance parameters. Each of example 700 and 800 is a bar plot with the importance metric (Mean absolute SHapley Additive exPlanations values, i.e., Mean |SHAP|) along the horizontal (x) axisAttorney Docket No. 17367-0160W01and lists configuration and material features along the vertical (y) axis, with error bars indicating variability of the importance estimates.[0001701 FIG. 9 shows an example 900 of the performance of the described techniques for single-unit organic light-emitting diodes, for tandem organic light-emitting diodes, and for single-unit perovskite-absorber based solar cells. In particular, FIG. 9 shows a bar chart titled “Comparison of Featurization,” with the vertical axis representing Test R2and the horizontal axis listing device types and performance parameter the Test R2is measured on (i.e., “OLED PE” refers the power efficiency performance parameter of single-unit organic light-emitting diodes; “Tandem-OLED PE” refers to the power efficiency performance parameter of tandem organic light-emitting diodes; and “Perovskite PCE” refers to the power conversion efficiency of singleunit perovskite-absorber based solar cells).[000171] The examples 400-600 of the performance of the described techniques in FIGS. 4-6 are evaluated in terms of performance metrics such as coefficient of determination R2using one or more datasets of real world device performance parameters for real world devices. Higher values of R2indicate the described techniques generate performance parameters(s) for devices that more closely align with their true performance parameter values. The R2values range from -co to 1, where 1 signifies the ML model predicts performance parameters perfectly and values below 0 signifies the ML model cannot predict performance parameters well at all. The predicted performance parameters include external quantum efficiency (EQE), current efficiency (CE), power efficiency (PE), electroluminescence peak position (EL(max)), electroluminescence bandwidth (EL(FWHM)), color coordinates (CIE(x), CIE(y)), and power conversion efficiency (PCE).[000172] The configurations of devices presented in FIGS. 4-9, may be described by the following string: [Anode | {Layer 1 | Layer 2 | ... | Layer N} | Cathode], where “{Layer x}“ represents the set of one or more chemical layers positioned between the electrodes and “|” represents the interface between adjacent layers. Note that layer labels are used for discussion below and are not formally required as descriptors by ML model 116.[000173] The configuration of OLED devices in FIGS. 4, 5, 7, and 9 contain one or more emission layers (EMLs), depending on if they adopt a single-unit or tandem architecture. As particular examples of other layer functionalities found in these device configurations include the hole injection layer (HIL), hole transport layer (HTL), electron injection layer (EIL), electronAttorney Docket No. 17367-0160W01transport layer (ETL), electron blocking layer (EBL), charge generation layer (CGL), interconnecting layer (ICL), encapsulation layer (ECL), and so on. There may be zero, one, or multiple of any such layer types in a given device.[000174] The configuration of a PSC device in FIGS. 6 and 8 contain one perovskite absorber layer (PAL). The other, non-electrode layers assume similar functionalities cited above, but are labeled according to their position relative to the electrode and absorber layers. Layers between the anode and the PAL are labeled as “ATL” and layers between the PAL and cathode are labeled “BTL”. This is to provide simple and unified descriptions of standard (n-i-p) and inverted (p-i-n) configurations where HTL(s) and ETL(s) positions are swapped. There may be zero, one, or multiple of any such layer types in a given device.[000175] FIG. 4 shows an example 400 of the performance of the described techniques for single-unit organic light emitting diodes. In particular, FIG. 4 shows a bar chart titled “Comparison of Featurization,” with the vertical axis representing TestR2over 1,826 single unit OLED devices and the horizontal axis listing device performance parameters (EQE, CE, PE, EL(max), EL(FWHM), CIE(x), and CIE(y)). The legend identifies three featurization conditions of the described techniques: Structure Only (hatched), Properties Only (light gray), and Structure+Properties (dark gray).[000176] For each column - “Structure Only”, “Properties Only”, and “Structure+Properties” - the device performance forecasting system processes both configuration input and material input. The difference in the three columns involves which descriptors are used to describe each layer material in the material input. For the “Structure Only” column, the material input descriptors are only composed of features that describe the chemical structure of the layer components, i.e., atoms, bonds, etc. For the “Properties Only” column, the material input descriptors are only composed of features that describe the physicochemical properties of the layer components, e.g., HOMO, LUMO, triplet energy, and so on. Finally, the “Structure+Properties” column, the material input utilizes a feature vector that is a concatenation of both the chemical structure descriptors and physicochemical property descriptors.[000177] Across the listed predicted performance parameters, the “Structure+Properties” bars are generally higher than the bars for either “Structure Only” or “Properties Only”, indicating higher Test R2when the device performance forecasting system jointly encodes andAttorney Docket No. 17367-0160W01processes configuration features, as well as material features composed of both structural features and property features. When implemented as described above, the device performance forecasting system improves predictive performance for EQE, CE, PE, EL(max), EL(FWHM), CIE(x), and CIE(y) relative to using material input composed of either only structural input parameters (e.g., SMILES) or only property input parameters (e.g., HOMO, LUMO).[000178] FIGS. 4, 7, and 9 depict performance prediction data for a dataset of 1,826 singleunit OLED devices obtained from published journal articles. The dataset consists of a diverse range of architectural configurations and constituent materials. The most common device configuration, adopted by 195 devices, involves: [Anode | HIL | EML | ETL | EIL | Cathode], The most common numbers of layers are eight and seven, each seen in over 450 devices, respectively. The smallest configuration consists of three layers, [Anode | EML | Cathode], and the largest configuration consists of eleven layers, [Anode | HIL | HTL | EBL | EBL | EML | ETL | ETL | EIL | Cathode], Most devices differ in the number and types of layers and / or the constituent materials. However, the dataset also exhibits cases of distinction between two devices, including (1) holding the overall configuration and chemical materials constant, while adjusting the emitter doping ratio, (2) holding the layer order and chemical materials constant, while adjusting the thickness of one or more layers, and so on.[000179] Regarding the distribution of materials in the dataset, there are over 1750 unique emitters. Most non-emitter layers (e.g., EBL, EIL, ETL, HBL, HIL, HTL) were fabricated using a set of approximately 20 materials. Most of the materials comprising all layers in the dataset are organic molecules or polymers. In addition, some non-emitter layers are composed of inorganic films although over 100 emitters correspond to organometallic complexes containing metal centers such as Pt, Ir, Zn, and Cu. The anode in all devices consists of indium tin oxide (ITO), while the cathode consists of a variety of metals, including Al, Ag, Ca, Ba, and Mg.[000180] FIG. 5 shows an example 500 of the performance of the described techniques for tandem organic light-emitting diodes. In particular, FIG. 5 shows a bar chart titled “Comparison of Featurization,” with the vertical axis representing Test R2over 184 OLED devices (87 singleunit and 97 tandem) and the horizontal axis listing device performance parameters (EQE and PE). The legend identifies three featurization conditions of the described techniques “Structure Only” (hatched), “Properties Only” (light gray), and “Structure+Properties” (dark gray), each having the same meaning as described above for FIG. 4.Attorney Docket No. 17367-0160W01[000181] For the EQE performance parameter, the “Structure+Properties” bar is at similar height to the “Structure Only” bar and higher than the “Properties Only” bar. For the PE performance parameter, the “Structure+Properties” bar is higher than the bars for either “Structure Only” or “Properties Only,”. Together, this indicates higher Test R2when the device performance forecasting system jointly encodes and processes configuration features as well as material features composed of both structural features and property features.[000182] FIGS. 5, and 9 depict performance prediction data for a dataset of devices consisting of 87 single-unit OLED devices and 97 tandem OLED devices obtained from a common set of published journal articles. By “tandem”, we refer to OLED devices with two or more emitters. The number of emitter layers in the tandem devices span 2-6. The number of devices in the dataset with 2, 3, 4, and 6 emitter layers is 83, 7, 6, and 1, respectively. The most common configuration of the tandem OLED devices, adopted by 195 devices, involves: [Anode | HIL | EML | ETL | EIL | Cathode], The most common numbers of layers are eight, which is seen in 500 devices, respectively. The smallest configuration consists of eight layers, [Anode | HTL | HIL | EML | CGL | EML | ETL | Cathode], and the largest configuration, consists of 20 layers, [Anode | HIL | HTL | HTL | EML | ETL | CGL | CGL | HTL | HTL | EML | ETL | CGL | CGL | HTL | HTL | EML | ETL | EIL | Cathode], Distinguishing cases for different device configurations are similar to those described for the single-unit OLED dataset. One key new example for this dataset involves varying the number of emitter layers and corresponding CGLs for a common configuration. Material distribution is similar to the single-unit OLEDs: most materials are organic molecules and polymers, with some amount of organometallic complexes and inorganic films.[000183] FIG. 6 shows an example 600 of the performance of the described techniques for single-unit perovskite-absorber based solar cells. In particular, FIG. 6 shows a scatter plot comparing Predicted PCE [%] (vertical axis) to Actual PCE [%] (horizontal axis) for perovskite solar cells (PSCs) devices, with markers on the bisectrix line indicating perfect prediction.Markers are shown for N = 164 devices, with darker markers labeled Train and lighter markers labeled Test; experimental error bars for the true real world performance parameter value are displayed for each marker. The example 700 shows Train R2= 0.82, Train RMSE (root mean squared error) = 2.20, Test R2= 0.79, and Test RMSE = 2.30.Attorney Docket No. 17367-0160W01[000184] FIGS. 6 and 8 depict performance prediction data for a dataset of 164 single-unit PSC devices. The most common configuration, adopted by 89 devices, involves: [Anode | ATL-1 | PAL | BTL-1 | Cathode], The most common numbers of layers are five and six, seen in over 85 devices 89, respectively. The smallest configuration consists of four layers, [Anode | ATL-1 | PAL | Cathode], and the largest configuration consists of eight layers, [Anode | ATL-1 | ATL-2 | PAL | BTL-1 | BTL-2 | BTL-3 | Cathode], The set contains approximately 126 standard (n-i-p) and 38 inverted (p-i-n) configurations. Most devices differ in the number of layers and / or constituent materials. However, the dataset also exhibits cases of distinction between two devices, including, (1) the overall configuration is held constant while the perovskite stoichiometry is adjusted, and (2) variably employing mixed use of both mesoporous films and compact films vs. solely compact films.[000185] In terms of materials, there are 90 unique perovskite layer formulations used to construct all of the devices. These formulations include 68 unique ABX3 crystal stoichiometries, as well as 20 different PAL additives used for passivation, stabilization, and so on. The set of materials that comprise ABX3 correspond to A={methylammonium, formamidinium, ethylammonium, ethylenediammonium, phenylethylammonium, K+, Cs+, Rb+}, B={Pb2+, Sn2+, In2+, Ge2+}, X={C1- Br-, I-}. These materials are formulated in different composition ratios. 60 materials represent the various ETL and HTL multilayers between the electrodes and the PAL. Of these materials, 24 are inorganic polycrystalline compounds (e.g., TiO2, SnO2, NiO, In2S3) and the remaining organic molecules and polymers arranged into amorphous films. The anodes consist of either indium tin oxide and fluorine-doped indium-tin oxide. The cathode materials range from Ag, Al, Au, C, Ca, and Cu.[000186] In FIG. 6, the device performance forecasting system can generate PCE predictions that align with real measured values for PSCs devices, as indicated by the clustering of Train and Test points near the bisectrix line and the reported R2and RMSE. FIG. 6 thus illustrates that the device performance forecasting system can produce accurate performance parameters for PSC devices, e.g., PCE, based on configuration input and material input for all layers of the device.[000187] As described above, FIG. 7 and FIG. 8 show examples 700 and 800 of configuration and material features with the largest importance metric values with regard to predicting device performance parameters. Each of example 700 and 800 is a bar plot with theAttorney Docket No. 17367-0160W01importance metric (Mean absolute SHapley Additive exPlanations values, i.e., Mean |SHAP|) along the horizontal (x) axis and lists configuration and material features along the vertical (y) axis, with error bars indicating variability of the importance estimates.[000188] That is, in some implementations, the described techniques (e.g., the device performance forecasting system) can determine importance metrics for configuration and material features to quantify each feature’s contribution to the accuracy of performance parameter(s) generated by the ML model. This allows the system to identify features that most influence predictions for a given device type.[000189] In particular, to determine an importance metric for each feature, the device performance forecasting system determines, for each feature, all possible subsets of configuration features and material features that do not include the feature. Then the system, for each feature and for each subset, uses the ML model to generate one or more performance parameters twice, once using the feature and the subset of features and once using just the subset of features. The system then determines, for each feature and for each subset, a marginal contribution for each performance parameter as the difference between performance parameter generated from processing the feature and the subset of features using the ML model and the respective performance parameter generated from processing just the subset of features using the ML model. Next, the system determines, for each feature and for each marginal contribution, a weight based on the size of the subset of features used to determine the marginal contribution. Finally, the system determines an importance metric, for each feature and for each performance parameter, as the sum of respective marginal contributions, each scaled by its respective weight.[000190] As an example, importance metric, for each feature, the system can determine the Shapley value of the feature according to the formula:where pLis the importance metric of feature i (i.e., the Shapley value of feature i) for ML model denoted as v, n denotes the total number of features, the summation is over each subset S of all possible subsets that do not include feature i (i.e., N \ {i}), and the termthefraction of times the subset S appears within the possible feature permutations; followingAttorney Docket No. 17367-0160W01Lundberg, Scott M., and Su-In Lee. “A unified approach to interpreting model predictions.” Advances in Neural Information Processing Systems (2017).[0001911 The system’s ability to compute importance metrics for each feature enables users to interpret how configuration and materials features influence multicomponent layered device performance parameters, supporting rational device design.[000192] FIG. 7 shows an example of importance metric values for certain configuration and material features of single-unit organic light-emitting diodes that are identified to be significant with regard to predicting the device performance parameter: external quantum efficiency (EQE).[000193] FIG. 8 shows an example of importance metric values for certain configuration and material features of single-unit perovskite-absorber based solar cells (PSCs) that are identified to be significant with regard to predicting the device performance parameter: power conversion efficiency (PCE).[000194] Examples 700 and 800 show the importance of features that correspond to structural and physicochemical properties of certain layers to predicting EQE in OLEDs and PCE in PSCs, respectively. The feature labels are formatted as ‘(Material Layer Group Type)-(Aggregation Type)-(Feature Name)’. Some example feature labels include the ‘Emitter-mean-Singlet-triplet_energy_gap,’ which is the mean of the singlet-triplet energy gaps for all of the materials classified as ‘Emitter’ in an OLED. Other example features are formatted as “MorganFingerprint_(number of bits)_(radius)_(index of the feature)” for characterization of molecular structures or “Perovskite-(Feature Name)” for characterization of certain structural components of perovskite cells, or “(Material Layer Group Type)-(Nanostructure Descriptor)” for characterization of overall film or nanostructure components.[000195] In particular, example 700 includes features corresponding to structural and property features for the emitter molecules and host material of the EML. Enumerated emitter molecule structural features include (1) “MorganFingerprint-674 3 112” and “MorganFingerprint-674_3_399”, which correspond to tri center boron fragments; (2) “MorganFingerprint-674_3_9”, which corresponds to benzyl fragments para-substituted with -NR and -CH(NR)2 groups; and (3) “MorganFingerprint-674_3_353”, which corresponds to dicyano and ethoxy substituted conjugated aliphatic fragment. Enumerated emitter molecule property features include (1) “Singlet-triplet_energy_gap” and “Triplet_Energy”, whichAttorney Docket No. 17367-0160W01correspond to the probability to fluorescence; and (2) “Hole reorganization energy”, which corresponds to charge mobility. Enumerated host material structural features include “MorganFingerprint-674 3 399”. Enumerated host material property features include “Electron_reorganization_energy”.[000196] Additionally, the example 800 includes features corresponding to the structural and property features for the cathode materials, perovskite materials in the PAL, and materials in the layers between the anode and the PAL (i.e., ATL). Enumerated cathode structural features include “MACCS_12”, which accounts for the presence of any of the following atoms: Cu, Zn, Ag, Cd, Au, or Hg. Enumerated cathode property features include “Scaled_LUMO” and “Hole Reorganization Energy”, which correspond to charge mobility. Enumerated ATL material structural features include: (1) “Morgan_693_4_485”, which corresponds to a planar conjugated tricycle fragment; and (2) “morphology”, which describes if the layer consists of an amorphous film or polycrystalline film composed of different nanostructures (e.g., nanoparticles, nanorods, nanowires, and so on). Enumerated ATL property features include“Triplet reorganization energy”, which corresponds to charge mobility. Enumerated perovskite material structural features include (1) "avg_dev Number", which measures how much the atomic numbers of input elements deviate from the composition’s average as a metric of a chemical diversity; (2) "avg_dev NUnfilled", which captures the electronic heterogeneity of the structure by measuring the variation in empty outer-shell electron orbitals across the ABX sites, making it a key indicator for predicting bandgap. Enumerated perovskite property features include "mean_MeltingT", which provides the weighted average of the melting temperatures of all atoms.[000197] Example 700 illustrates how features that correspond to both structural descriptors and chemical properties of the emitter material(s) layers and host material layers significantly impact the external quantum efficiency (EQE) of the OLED devices. The features corresponding to structural descriptors are highlighted by the “MorganFingerprint_*” features; whereas the features correspond to relevant chemical properties include the singlet-triplet gap and hole reorganization energy of the emitter and the electron reorganization energy of the host.[000198] Example 800 illustrates how features that correspond to both structural descriptors and chemical properties of perovskites and as well as surrounding layers, such as the cathode and “ATL” layers significantly impact the power conversion efficiency (PCE) of the PSC devices.Attorney Docket No. 17367-0160W01Here, “ATL” refers to one or more transport layers between the anode and perovskite layers. Significant structural features include “Perovskite avg_dev_Number” (the average absolute deviation of the number of unfilled orbitals perovskite absorber in a PSC) and “ATL_morphology” which is an encoding of whether constituent transport layers are composed of nanoparticles, nanorods, nanobundles, and so on. Significant chemical features span multiple layer types as well, including the LUMO energy levels and triplet reorganization energies.[000199] Examples 700 and 800 demonstrate how the described techniques’ determination of an importance metric for each feature can improve understanding of the relationship between material, property and device importance. It is evident that a combination of structural and property features of multiple components of multiple layers is significant to making accurate predictions.[000200] As described above, FIG. 9 shows an example 900 of the performance of the described techniques for single-unit organic light-emitting diodes, for tandem organic lightemitting diodes, and for single-unit perovskite-absorber based solar cells. In particular, FIG. 9 shows a bar chart titled “Comparison of Featurization,” with the vertical axis representing Test R2and the horizontal axis listing device types and performance parameter the Test R2is measured on (i.e., “OLED PE” refers the power efficiency performance parameter of single-unit organic light-emitting diodes; “Tandem-OLED PE” refers to the power efficiency performance parameter of tandem organic light-emitting diodes; and “Perovskite PCE” refers to the power conversion efficiency of single-unit perovskite-absorber based solar cells).[000201] The legend identifies two featurization conditions: “Critical Layer(s) Only” (hatched) and “Full Device” (dark gray). The “Critical Layer(s) Only” columns correspond to the device performance forecasting system processing configuration features and material features from one or more layers identified as critical (e.g., emissive for OLEDs, absorber for PSCs), while the “Full Device” columns correspond to the device performance forecasting system processing configuration features and material features for all layers of the device.[000202] Across the plotted properties, the “Full Device” bars are higher than the “Critical Layer(s) Only” bars, indicating higher Test R2when the device performance forecasting system encodes and processes features from all layers. Thus, example 900 demonstrates the increased performance of models when featurizing and processing all layers of the device versus certain ‘critical’ layers (where the principal physicochemical mechanisms of action behind deviceAttorney Docket No. 17367-0160W01operation occur, e.g., the emission layer for single-unit and tandem organic light-emitting diodes and absorber layer for perovskite-absorber based solar cells).[0002031 FIG. 10 shows an example of user-specified data interfacing with a device performance forecasting system, e.g., the device performance forecasting system 100 of FIG. 1, appropriately programmed in accordance with this specification.[000204] In some implementations, the system 100 can be a part of a user-driven customization of the ML model 116. The system 100 facilitates fine-tuning or retraining of the ML model 116 using user-specified data 1000 (i.e., data that includes user-specified performance parameters to be generated by the ML model 116, user-specified input features, and datasets that include configuration and material input database 1002 of a plurality of devices), allowing the user to modify the ML model 116 training objectives and adapt its predictive capabilities to specific applications. By further training the ML model 116, e.g., further training the ML model 116 as was described with reference to FIG. 3 above, the system 100 ensures accuracy and relevance of predictions for user-specific contexts.[000205] In particular, the system 100 can adjust the trainable parameters of the ML model 116 based on user-specified data 1000, without modifying the ML model’s 116 configuration, e.g., neural network architecture or general hyperparameters, and generate one or more user-specified performance parameters 104 using the ML model 116.[000206] Such a capability of the system 100 allows the ML model 116 to remain flexible to the needs of a user, and a user can readily update or add new material(s) or multicomponent layered device features and explain how ML model predictions change, thus enhancing the knowledge of device architecture or related chemical properties.[000207] For example, while the ML model 116 originally trained by the system 100 makes accurate predictions composed of a variety of materials, layer compositions, and layer ordering, a user may desire the diversity of the devices to be localized to a part of the device undergoing investigation by the user.[000208] As a particular example, a user may only be interested in exploring the component or dopant percentage of a single layer with other layers being fixed. As another particular example, a user may only be interested in exploring adding a new layer to a certain part of the device while maintaining size, order, and chemical composition of each layer. The system 100Attorney Docket No. 17367-0160W01can obtain training data from the user aligned with either of these goals, train the ML model 116 using the data, and provide the user accurate performance parameters.[0002091 Insome cases, the systems described above can be used as part of an industrial design process, e.g., as a tool that predicts a device’s performance parameters. For example, after the system 100 trains the ML model 116, the system 100 can be incorporated as a component of a larger composable process or system, where the larger process or system utilizes the system 100’s capability to receive input characterizing a device (e.g., the configuration input 102A and material input 102B) and predict values for performance parameters(s) 104 of the device, e.g., which may then be compared to target design criteria of the device being designed.[000210] As one example of an industrial design process, the system 100 is used in a large-scale device screening process in which, a designer screens a library of device prototypes by using the system 100 to predict the performance parameters) 104 of each device prototype and then select one or more device prototypes to further evaluate based on the predicted performance param eter(s) 104 of each.[000211] As another example, the system 100 can be part of a multi -parameter optimization process. That is, a computer program or algorithm that varies device inputs (i.e., configuration input 102A and material input 102B) to fit design goals for a new device can use the system 100 to determine the performance param eter(s) associated with chosen configuration input 102A and material input 102B and then vary the device inputs accordingly. The multi -parameter optimization process can be one of a variety of optimization processes, e.g., gradient-based methods, evolutionary algorithms, Bayesian optimization, and so on.[000212] Ultimately, devices based on the designs resulting from these processes can be fabricated, for experimental testing and / or mass production.[000213] Using the system 100 as part of an industrial design and industrial device production process can be efficient. For example, using the system 100 to predict performance of device designs can reduce the amount of experimental work needed to optimize a device’s design and the associated time and expense.[000214] This specification uses the term “configured” in connection with systems and computer program components, such as computer system 1100 described below.[000215] FIG. 11 shows an example computer system 1100. The example computer system 1100 is one in which embodiments of the present disclosure may be implemented.Attorney Docket No. 17367-0160W01[000216] For a system of one or more computers to be configured to perform particular operations or actions means that the system 1100 has installed on it: software, firmware, hardware (e.g., processor 1104 coupled to bus 1102), or a combination of them that in operation cause the system 1100 to perform the operations or actions. For one or more computer programs to be configured to perform particular operations or actions means that the one or more programs include instructions that, when executed by data processing apparatus, such as processor 1104, cause the apparatus to perform the operations or actions. Embodiments of the subject matter and the functional operations described in this specification (e.g., processor 1104, main memory 1106, storage device 1108, I / O interface 1110, and communication interface 1116) can be implemented in digital electronic circuitry, in tangibly embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non transitory storage medium for execution by, or to control the operation of, data processing apparatus (e.g., processor 1104). The computer storage medium can be a machine-readable storage device 1108, a machine-readable storage substrate, a random or serial access memory device (such as main memory 1106), or a combination of one or more of them. The computer system 1100 may further include input device 1112 (e.g., keyboards, sensor) and output devices 1114 (e g., displays, actuators) connected via the VO interface 1110. Alternatively, or in addition, the program instructions can be encoded on an artificially generated propagated signal, e.g., a machinegenerated electrical, optical, or electromagnetic signal, which is generated to encode information for transmission (e.g., via communication interface 1116 over network link 1118) to suitable receiver apparatus for execution by a data processing apparatus.[000217] The term “data processing apparatus” refers to data processing hardware and encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can also be, or further include, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit). The apparatus can optionally include, in addition to hardware, code that creates an execution environment forAttorney Docket No. 17367-0160W01computer programs, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them.[0002181 A computer program, which may also be referred to or described as a program, software, a software application, an app, a module, a software module, a script, or code, can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages; and it can be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data, e.g., one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, sub programs, or portions of code. A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a data communication network.[000219] In this specification, the term “database” is used broadly to refer to any collection of data: the data does not need to be structured in any particular way, or structured at all, and it can be stored on storage devices in one or more locations. Thus, for example, the index database can include multiple collections of data, each of which may be organized and accessed differently.[000220] Similarly, in this specification the term “engine” is used broadly to refer to a software-based system, subsystem, or process that is programmed to perform one or more specific functions. Generally, an engine will be implemented as one or more software modules or components, installed on one or more computers in one or more locations. In some cases, one or more computers will be dedicated to a particular engine; in other cases, multiple engines can be installed and running on the same computer or computers.[000221] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by special purpose logic circuitry, e.g., an FPGA or an ASIC, or by a combination of special purpose logic circuitry and one or more programmed computers.Attorney Docket No. 17367-0160W01[000222] Computers suitable for the execution of a computer program can be based on general or special purpose microprocessors or both, or any other kind of central processing unit. Generally, a central processing unit will receive instructions and data from a read only memory or a random access memory or both. The essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory devices for storing instructions and data. The central processing unit and the memory can be supplemented by, or incorporated in, special purpose logic circuitry. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device, e g., a universal serial bus (USB) flash drive, to name just a few.[000223] Computer readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto optical disks; and CD ROM and DVD-ROM disks.[000224] To provide for interaction with a user, embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user and a keyboard and a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user’s device in response to requests received from the web browser. Also, a computer can interact with a user by sending text messages or other forms of message to a personal device, e.g., a smartphone that is running a messaging application, and receiving responsive messages from the user in return.Attorney Docket No. 17367-0160W01[000225] Data processing apparatus for implementing ML models can also include, for example, special-purpose hardware accelerator units for processing common and computeintensive parts of machine learning training or production, i.e., inference, workloads.[000226] ML models can be implemented and deployed using a machine learning framework, e.g., a TensorFlow framework or a Jax framework.[000227] Embodiments of the subject matter described in this specification can be implemented in a computing system that includes a back end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front end component, e.g., a client computer having a graphical user interface, a web browser, or an app through which a user can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN) and a wide area network (WAN), e.g., the Internet.[000228] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In some embodiments, a server transmits data, e.g., an HTML page, to a user device, e.g., for purposes of displaying data to and receiving user input from a user interacting with the device, which acts as a client. Data generated at the user device, e.g., a result of the user interaction, can be received at the server from the device.[000229] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment.Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially be claimed as such, one or more features from a claimed combination can in some cases be excisedAttorney Docket No. 17367-0160W01from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.[0002301 Similarly, while operations are corresponded to in the drawings and recited in the claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.EMBODIMENTS[000231] Although the present invention is defined in the claims, it should be understood that the present invention can also (alternatively) be defined in accordance with the following embodiments:1. A method of using a machine learning (ML) model to predict performance parameters of a device comprising two or more layers, the method comprising:receiving, at a computer, configuration input for the device, wherein the configuration input comprises device-wide parameters and / or layer parameters for each of the two or more layers;receiving, at the computer, material input for the device, wherein the material input comprises composition values of each of the two or more layers;extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model, the configuration features comprising values derived from the configuration input formatted for processing by the ML model;extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model, the material features comprising values derived from the material input formatted for processing by the ML model; andAttorney Docket No. 17367-0160W01processing the configuration features and the material features using the ML model to generate one or more performance parameters, the ML model having been trained using a training dataset comprising a plurality of training examples, each training example comprising values for one or more target performance parameters and respective configuration features and material features, wherein the ML model is trained to process the respective configuration features and respective material features and output a value for each of the one or more performance parameters.2. The method of embodiment 1, wherein training the ML model comprises:obtaining the training dataset comprising training examples;for each training example, processing the respective configuration features and materials features and generating one or more performance parameters;evaluating an objective function using the one or more performance parameters for each training example; andupdating the ML model trainable parameters to optimize the objective function.3. The method of embodiment 1 or 2, wherein the composition values of each of the two or more layers comprises one or more chemical structures or one or more chemical properties.4. The method of embodiment 3, wherein one or more of the chemical properties are experimental measurements.5. The method of embodiment 3, wherein one or more of the chemical properties are computed electronic properties.6. The method of embodiment 3, wherein one or more of the chemical properties are computed ensemble and / or morphological properties.7. The method of embodiment 3, wherein one or more of the chemical properties are computed using pre-trained surrogate models.Attorney Docket No. 17367-0160W018. The method of embodiment 3, wherein one or more chemical properties are as described in any one of embodiments 4 to 7.9. The method of embodiment 3, wherein one or more chemical properties explicitly encode the physical and chemical interactions and / or relationships between layers.10. The method of any one of embodiments 1 to 9, wherein the configuration input further comprises the overall device-wide structural and operational parameters, as well as the layerspecific parameters of dopant composition ratios and layer fabrication methods for each of the two or more layers.11. The method of any one of embodiments 1 to 10, further comprising maintaining data of material input values, and wherein extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model comprises:determining whether the material input is missing one or more values;retrieving the one or more missing values from maintained data; andcomputing the one or more missing values, wherein the computing is performed for any missing values not retrieved from the maintained data.12. The method of embodiment 10, wherein computing the missing values of the material input comprises using a pre-trained surrogate model.13. The method of any one of embodiments 1 to 12, wherein extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model comprises:determining whether the configuration input is missing one or more values; and in response to determining that the configuration input is missing one or more values, imputing configuration input missing values.Attorney Docket No. 17367-0160W0114. The method of embodiment 12, wherein imputing the configuration input missing values comprises using a pre-trained imputer neural network.15. The method of any one of embodiments 1 to 14, further comprising:determining, for each feature of the configuration features and material features, all possible subsets of configuration features and material features that do not include the feature;processing, for each feature and for each subset, the feature and the subset of features using the ML model to generate one or more performance parameters;processing, for each feature and for each subset, the subset of features using the ML model to generate one or more performance parameters;determining, for each feature and for each subset, a marginal contribution for each performance parameter as the difference between performance parameter generated from processing the feature and the subset of features using the ML model and the respective performance parameter generated from processing the subset of features using the ML model; determining, for each feature and for each marginal contribution, a weight based on the size of the subset of features used to determine the marginal contribution; anddetermining an importance metric, which quantifies the significance of each feature with regard to predictions of each performance parameter, as the sum of respective marginal contributions, each scaled by its respective weight.16. The method of any one of embodiments 1 to 15, further comprising:adjusting the trainable parameters of the ML model based on user-specified data, without modifying the ML model’s configuration; and generating one or more user-specified performance parameters using the ML model.17. The method of any one of embodiments 1 to 16, wherein the two or more layers are layers of a device selected from the group consisting of a light emitting diode, transistor, a photovoltaic cell, a capacitor, a battery, a sensor, a flexible display, and a multilayer printed circuit board.Attorney Docket No. 17367-0160W0118. The method of any one of embodiments 1 to 17, where the ML model is a model selected from the group consisting of an artificial neural network, a tree-based model, a support vector machine, a k-nearest neighbor model, and a generalized linear model.19. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one more computers to perform operations using a machine learning (ML) model to predict performance parameters of a device comprising two or more layers comprising:receiving, at a computer, configuration input for the device, wherein the configuration input comprises device-wide parameters and layer parameters for each of the two or more layers;receiving, at the computer, material input for the device, wherein the material input comprises composition values of each of the two or more layers;extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model, the configuration features comprising values derived from the configuration input formatted for processing by the ML model;extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model, the material features comprising values derived from the material input formatted for processing by the ML model; and processing the configuration features and the material features using the ML model to generate one or more performance parameters, the ML model having been trained using a training dataset comprising a plurality of training examples, each training example comprising values for one or more target performance parameters and respective configuration features and material features, wherein the ML model is trained to process the respective configuration features and respective material features and output a value for each of the one or more performance parameters.20. The system of embodiment 19, wherein training the ML model comprises:obtaining the training dataset comprising training examples;for each training example, processing the respective configuration features and materials features and generating one or more performance parameters;Attorney Docket No. 17367-0160W01evaluating an objective function using the one or more performance parameters for each training example; andupdating the ML model trainable parameters to optimize the objective function.21. The system of embodiment 19 or 20, wherein the composition values of each of the two or more layers comprises one or more chemical structures or one or more chemical properties.22. The system of embodiment 21, wherein one or more of the chemical properties are experimental measurements.23. The system of embodiment 21, wherein one or more of the chemical properties are computed electronic properties.24. The system of embodiment 21, wherein one or more of the chemical properties are computed ensemble and / or morphological properties.25. The system of embodiment 21, wherein one or more of the chemical properties are computed using pre-trained surrogate models.26. The system of embodiment 21, wherein one or more chemical properties are as described in any one of embodiments 22 to 25.27. The system of embodiment 21, wherein one or more chemical properties explicitly encode the physical and chemical interactions and / or relationships between layers.28. The system of any one of embodiments 19 to 27, wherein the configuration input further comprises the overall device-wide structural and operational parameters, as well as the layerspecific parameters of dopant composition ratios and layer fabrication methods for each of the two or more layers.Attorney Docket No. 17367-0160W0129. The system of any one of embodiments 19 to 28, further comprising maintaining data of material input values, and wherein extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model comprises:determining whether the material input is missing one or more values;retrieving the one or more missing values from maintained data; andcomputing the one or more missing values, wherein the computing is performed for any missing values not retrieved from the maintained data.30. The system of embodiment 28, wherein computing the missing values of the material input comprises using a pre-trained surrogate model.31. The system of any one of embodiments 19 to 30, wherein extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model comprises:determining whether the configuration input is missing one or more values; and in response to determining that the configuration input is missing one or more values, imputing configuration input missing values.32. The system of embodiment 30, wherein imputing the configuration input missing values comprises using a pre-trained imputer neural network.33. The system of any one of embodiments 19 to 32, further comprising:determining, for each feature of the configuration features and material features, all possible subsets of configuration features and material features that do not include the feature;processing, for each feature and for each subset, the feature and the subset of features using the ML model to generate one or more performance parameters;processing, for each feature and for each subset, the subset of features using the ML model to generate one or more performance parameters;determining, for each feature and for each subset, a marginal contribution for each performance parameter as the difference between performance parameter generated fromAttorney Docket No. 17367-0160W01processing the feature and the subset of features using the ML model and the respective performance parameter generated from processing the subset of features using the ML model; determining, for each feature and for each marginal contribution, a weight based on the size of the subset of features used to determine the marginal contribution; anddetermining an importance metric, which quantifies the significance of each feature with regard to predictions of each performance parameter, as the sum of respective marginal contributions, each scaled by its respective weight.34. The system of any one of embodiments 19 to 33, further comprising:adjusting the trainable parameters of the ML model based on user-specified data, without modifying the ML model’s configuration; and generating one or more user-specified performance parameters using the ML model.35. The system of any one of embodiments 19 to 34, wherein the two or more layers are layers of a device selected from the group consisting of a light emitting diode, transistor, a photovoltaic cell, a capacitor, a battery, a sensor, a flexible display, and a multilayer printed circuit board.36. The system of any one of embodiments 19 to 35, where the ML model is a model selected from the group consisting of an artificial neural network, a tree-based model, a support vector machine, a k-nearest neighbor model, and a generalized linear model.37. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one more computers to perform operations using a machine learning (ML) model to predict performance parameters of a device comprising two or more layers comprising:receiving, at a computer, configuration input for the device, wherein the configuration input comprises device-wide parameters and layer parameters for each of the two or more layers;receiving, at the computer, material input for the device, wherein the material input comprises composition values of each of the two or more layers;Attorney Docket No. 17367-0160W01extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model, the configuration features comprising values derived from the configuration input formatted for processing by the ML model;extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model, the material features comprising values derived from the material input formatted for processing by the ML model; and processing the configuration features and the material features using the ML model to generate one or more performance parameters, the ML model having been trained using a training dataset comprising a plurality of training examples, each training example comprising values for one or more target performance parameters and respective configuration features and material features, wherein the ML model is trained to process the respective configuration features and respective material features and output a value for each of the one or more performance parameters.38. The one or more non -transitory computer storage media of embodiment 37, wherein training the ML model comprises:obtaining the training dataset comprising training examples;for each training example, processing the respective configuration features and materials features and generating one or more performance parameters;evaluating an objective function using the one or more performance parameters for each training example; andupdating the ML model trainable parameters to optimize the objective function.39. The one or more non-transitory computer storage media of embodiment 37 or 38, wherein the composition values of each of the two or more layers comprises one or more chemical structures or one or more chemical properties.40. The one or more non-transitory computer storage media of embodiment 39, wherein one or more of the chemical properties are experimental measurements.Attorney Docket No. 17367-0160W0141. The one or more non-transitory computer storage media of embodiment 39, wherein one or more of the chemical properties are computed electronic properties.42. The one or more non-transitory computer storage media of embodiment 39, wherein one or more of the chemical properties are computed ensemble and / or morphological properties.43. The one or more non-transitory computer storage media of embodiment 39, wherein one or more of the chemical properties are computed using pre-trained surrogate models.44. The one or more non-transitory computer storage media of embodiment 39, wherein one or more chemical properties are as described in any one of embodiments 40 to 43.45. The one or more non-transitory computer storage media of embodiment 39, wherein one or more chemical properties explicitly encode the physical and chemical interactions and / or relationships between layers.46. The one or more non-transitory computer storage media of any one of embodiments 37 to 45, wherein the configuration input further comprises the overall device-wide structural and operational parameters, as well as the layer-specific parameters of dopant composition ratios and layer fabrication methods for each of the two or more layers.47. The one or more non-transitory computer storage media of any one of embodiments 37 to 46, further comprising maintaining data of material input values, and wherein extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model comprises:determining whether the material input is missing one or more values;retrieving the one or more missing values from maintained data; andcomputing the one or more missing values, wherein the computing is performed for any missing values not retrieved from the maintained data.Attorney Docket No. 17367-0160W0148. The one or more non-transitory computer storage media of embodiment 46, wherein computing the missing values of the material input comprises using a pre-trained surrogate model.49. The one or more non-transitory computer storage media of any one of embodiments 37 to 48, wherein extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model comprises:determining whether the configuration input is missing one or more values; and in response to determining that the configuration input is missing one or more values, imputing configuration input missing values.50. The one or more non-transitory computer storage media of embodiment 48, wherein imputing the configuration input missing values comprises using a pre-trained imputer neural network.51. The one or more non-transitory computer storage media of any one of embodiments 37 to 50, further comprising:determining, for each feature of the configuration features and material features, all possible subsets of configuration features and material features that do not include the feature;processing, for each feature and for each subset, the feature and the subset of features using the ML model to generate one or more performance parameters;processing, for each feature and for each subset, the subset of features using the ML model to generate one or more performance parameters;determining, for each feature and for each subset, a marginal contribution for each performance parameter as the difference between performance parameter generated from processing the feature and the subset of features using the ML model and the respective performance parameter generated from processing the subset of features using the ML model; determining, for each feature and for each marginal contribution, a weight based on the size of the subset of features used to determine the marginal contribution; andAttorney Docket No. 17367-0160W01determining an importance metric, which quantifies the significance of each feature with regard to predictions of each performance parameter, as the sum of respective marginal contributions, each scaled by its respective weight.52. The one or more non-transitory computer storage media of any one of embodiments 37 to 51, further comprising:adjusting the trainable parameters of the ML model based on user-specified data, without modifying the ML model’s configuration; and generating one or more user-specified performance parameters using the ML model.53. The one or more non-transitory computer storage media of any one of embodiments 37 to 52, wherein the two or more layers are layers of a device selected from the group consisting of a light emitting diode, transistor, a photovoltaic cell, a capacitor, a battery, a sensor, a flexible display, and a multilayer printed circuit board.54. The one or more non-transitory computer storage media of any one of embodiments 37 to 53, where the ML model is a model selected from the group consisting of an artificial neural network, a tree-based model, a support vector machine, a k-nearest neighbor model, and a generalized linear model.[000232] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes corresponded to in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous.[000233] What is claimed is:

Claims

Attorney Docket No. 17367-0160W01CLAIMS1. A method of using a machine learning (ML) model to predict performance parameters of a device comprising two or more layers, the method comprising:receiving, at a computer, configuration input for the device, wherein the configuration input comprises device-wide parameters and / or layer parameters for each of the two or more layers;receiving, at the computer, material input for the device, wherein the material input comprises composition values of each of the two or more layers;extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model, the configuration features comprising values derived from the configuration input formatted for processing by the ML model;extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model, the material features comprising values derived from the material input formatted for processing by the ML model; and processing the configuration features and the material features using the ML model to generate one or more performance parameters, the ML model having been trained using a training dataset comprising a plurality of training examples, each training example comprising values for one or more target performance parameters and respective configuration features and material features, wherein the ML model is trained to process the respective configuration features and respective material features and output a value for each of the one or more performance parameters.

2. The method of claim 1, wherein training the ML model comprises:obtaining the training dataset comprising training examples;for each training example, processing the respective configuration features and materials features and generating one or more performance parameters;evaluating an objective function using the one or more performance parameters for each training example; andupdating the ML model trainable parameters to optimize the objective function.Attorney Docket No. 17367-0160W013. The method of claim 1, wherein the composition values of each of the two or more layers comprises one or more chemical structures or one or more chemical properties.

4. The method of claim 3, wherein one or more of the chemical properties are experimental measurements.

5. The method of claim 3, wherein one or more of the chemical properties are computed electronic properties.

6. The method of claim 3, wherein one or more of the chemical properties are computed ensemble and / or morphological properties.

7. The method of claim 3, wherein one or more of the chemical properties are computed using pre-trained surrogate models.

8. The method of claim 3, wherein one or more chemical properties are experimental measurements, computed electronic properties, computed ensemble and / or morphological properties, and properties computed using pre-trained surrogate models.

9. The method of claim 3, wherein one or more chemical properties explicitly encode the physical and chemical interactions and / or relationships between layers.

10. The method of claim 1, wherein the configuration input further comprises the overall device-wide structural and operational parameters, as well as the layer-specific parameters of dopant composition ratios and layer fabrication methods for each of the two or more layers.

11. The method of claim 1, further comprising maintaining data of material input values, and wherein extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model comprises:determining whether the material input is missing one or more values;retrieving the one or more missing values from maintained data; andAttorney Docket No. 17367-0160W01computing the one or more missing values, wherein the computing is performed for any missing values not retrieved from the maintained data.

12. The method of claim 10, wherein computing the missing values of the material input comprises using a pre-trained surrogate model.

13. The method of claim 1, wherein extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model comprises:determining whether the configuration input is missing one or more values; and in response to determining that the configuration input is missing one or more values, imputing configuration input missing values.

14. The method of claim 12, wherein imputing the configuration input missing values comprises using a pre-trained imputer neural network.

15. The method of claim 1, further comprising:determining, for each feature of the configuration features and material features, all possible subsets of configuration features and material features that do not include the feature;processing, for each feature and for each subset, the feature and the subset of features using the ML model to generate one or more performance parameters;processing, for each feature and for each subset, the subset of features using the ML model to generate one or more performance parameters;determining, for each feature and for each subset, a marginal contribution for each performance parameter as the difference between performance parameter generated from processing the feature and the subset of features using the ML model and the respective performance parameter generated from processing the subset of features using the ML model;determining, for each feature and for each marginal contribution, a weight based on the size of the subset of features used to determine the marginal contribution; anddetermining an importance metric, which quantifies the significance of each feature with regard to predictions of each performance parameter, as the sum of respective marginalAttorney Docket No. 17367-0160W01contributions, each scaled by its respective weight.

16. The method of claim 1, further comprising:adjusting the trainable parameters of the ML model based on user-specified data, without modifying the ML model’s configuration; and generating one or more user-specified performance parameters using the ML model.

17. The method of claim 1, wherein the two or more layers are layers of a device selected from the group consisting of a light emitting diode, transistor, a photovoltaic cell, a capacitor, a battery, a sensor, a flexible display, and a multilayer printed circuit board.

18. The method of claim 1, where the ML model is a model selected from the group consisting of an artificial neural network, a tree-based model, a support vector machine, a k-nearest neighbor model, and a generalized linear model.

19. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one more computers to perform operations using a machine learning (ML) model to predict performance parameters of a device comprising two or more layers comprising:receiving, at a computer, configuration input for the device, wherein the configuration input comprises device-wide parameters and layer parameters for each of the two or more layers;receiving, at the computer, material input for the device, wherein the material input comprises composition values of each of the two or more layers;extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model, the configuration features comprising values derived from the configuration input formatted for processing by the ML model;extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model, the material features comprising values derived from the material input formatted for processing by the ML model; andAttorney Docket No. 17367-0160W01processing the configuration features and the material features using the ML model to generate one or more performance parameters, the ML model having been trained using a training dataset comprising a plurality of training examples, each training example comprising values for one or more target performance parameters and respective configuration features and material features, wherein the ML model is trained to process the respective configuration features and respective material features and output a value for each of the one or more performance parameters.

20. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one more computers to perform operations using a machine learning (ML) model to predict performance parameters of a device comprising two or more layers comprising:receiving, at a computer, configuration input for the device, wherein the configuration input comprises device-wide parameters and layer parameters for each of the two or more layers;receiving, at the computer, material input for the device, wherein the material input comprises composition values of each of the two or more layers;extracting, by the computer and from the configuration input, features from the configuration input to generate configuration features for inputting into the ML model, the configuration features comprising values derived from the configuration input formatted for processing by the ML model;extracting, by the computer and from the material input, features from the material input to generate material features for inputting into the ML model, the material features comprising values derived from the material input formatted for processing by the ML model; and processing the configuration features and the material features using the ML model to generate one or more performance parameters, the ML model having been trained using a training dataset comprising a plurality of training examples, each training example comprising values for one or more target performance parameters and respective configuration features and material features, wherein the ML model is trained to process the respective configuration features and respective material features and output a value for each of the one or more performance parameters.

Images