Training method and electronic device suitable for density functional recommendation model

Abstract

The purpose of this invention is to provide a training method and electronic device suitable for density functional recommendation models. Through an active learning mechanism, the system only needs to perform expensive gold standard and multiple DFA calculations on a small portion of all molecules to efficiently train a model with comparable performance. This reduces the total computational cost of the model training phase, enabling the development of high-performance DFA recommendation systems to overcome economic bottlenecks.

Description

Technical Field

[0001] This invention relates to the field of computational chemistry, specifically to the optimization of computational cost in the training process of a DFA recommendation model. Background Technology

[0002] In computational chemistry and materials science, density functional theory (DFA) is a core tool for electronic structure calculations and virtual high-throughput screening. Its computational accuracy critically depends on the chosen exchange-correlation functional approximation. However, as a key parameter in this computational process, the DFA exhibits a significant systematic dependence on accuracy among different DFAs. While it may perform well for some chemical systems, it can introduce substantial errors for others.

[0003] In large-scale explorations across a broad chemical space, using a single DFA would not guarantee the reliability of calculation results for all research subjects; if the optimal DFA were empirically screened for each system using high-precision benchmark methods, the computational cost would become prohibitive.

[0004] To address the aforementioned issues, existing technologies propose a machine learning-based intelligent DFA recommendation method. This approach primarily consists of two phases: In the training phase, a training set containing a large number of representative molecules is pre-constructed, and two expensive computations are performed for each molecule: one is to calculate a reference value for the target attribute using a high-precision computational method, the so-called "gold standard" method; the other is to use a pool containing multiple candidate DFAs, for example, 48, to calculate the same attribute for each. By comparison, the absolute value of the true error for each DFA corresponding to each molecule is obtained, serving as the training label. Subsequently, a specific neural network model is independently trained for each DFA in the pool; 48 DFAs result in 48 neural network models. The learning objective is to predict the error of a DFA based on the molecular electron density features extracted from a single low-cost DFT calculation. In the application phase, for a new system, only one low-cost calculation is needed to obtain its electron density features, which are then input into all trained models. This allows for the parallel acquisition of the prediction errors of each DFA, ultimately recommending the DFA with the smallest prediction error.

[0005] While the existing technical solutions described above can select the optimal DFA, they have certain technical drawbacks, namely, the data acquisition cost during the model training phase is extremely high.

[0006] Specifically, in order to train a reliable model, high-quality supervision signals must be obtained for a large number of molecules in the training set. This requires performing a gold standard calculation and multiple DFA calculations for each molecule. The gold standard calculation itself is already extremely time-consuming and labor-intensive, and combining it with large-scale DFA calculations makes the overall computational cost of building the training set prohibitively high. Summary of the Invention

[0007] The purpose of this invention is to provide a training method and electronic device suitable for density functional recommendation models. Through an active learning mechanism, the system only needs to perform expensive gold standard and multiple DFA calculations on a small portion of all molecules to efficiently train a model with comparable performance. This reduces the total computational cost of the model training phase, enabling the development of high-performance DFA recommendation systems to overcome economic bottlenecks.

[0008] In a first aspect, the present invention provides a coordination design method based on a full history data retention strategy, comprising the following steps: collecting all historical samples and constructing an unsuccessful sample pool from the unsuccessful historical samples;

[0009] All samples in the non-successful sample pool are identified for failure reasons and divided into structural failure samples and performance failure samples. The structural failure samples are those that do not conform to the reasonable rules of chemical structure. The performance failure samples are those whose target attribute value P is lower than the preset target attribute threshold T_target.

[0010] A performance correlation model is constructed based on the aforementioned performance failure samples, and design strategies are provided for the design of new coordination compounds.

[0011] In a second aspect, the present invention provides an electronic device, including a processor and a memory; the processor is connected to the memory.

[0012] The memory is used to store executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to execute the method described above.

[0013] In summary, the present invention has the following beneficial effects:

[0014] 1. By introducing an active learning loop, only the initial small number of representative molecules and the high-uncertainty molecules intelligently screened in subsequent iterations require high-cost "gold standard" and multiple DFA calculations. This reduces the total training computation cost required to build a high-performance DFA error prediction model set, solving the core defect of traditional methods that are difficult to promote and apply due to the extremely high cost of acquiring training data.

[0015] 2. By employing a clustering algorithm to screen initial molecules based on physicochemical descriptors, we ensured that the initial training subset possessed sufficient diversity and representativeness in the chemical space. This avoided the model prematurely falling into local optima due to initial sample bias, thereby improving the robustness of the entire training process and the generalization ability of the final model.

[0016] 3. The mean squared error loss function is used in model training, which imposes a greater penalty on large error samples, forcing the model to prioritize optimizing the least accurate predictions, thus improving the model's reliability in the worst case. Combined with the adaptive learning rate and momentum characteristics of the Adam optimizer, this ensures that the training process can converge quickly and stably to a better solution, improving training efficiency and the accuracy of the final model.

[0017] 4. The statistical variance of N predicted values ​​is used as a specific measure of uncertainty. A large variance value directly indicates that the model group has a low consensus and poor confidence in the error prediction of a certain molecule. This enables efficient and accurate screening of the target molecules with the most training value and the greatest ability to improve model performance from a large number of unlabeled molecules.

[0018] 5. Set performance- or resource-based iteration termination conditions so that the training process can automatically stop when the accuracy requirements are met, or terminate in a timely manner when computing resources are exhausted. This avoids meaningless over-iteration or inappropriate waste of resources, ensuring that the method can output the current optimal model under any constraint, thus achieving a balance between cost and performance.

[0019] 6. Evaluate the actual recommendation performance of the model set on an independent validation set, i.e., calculate the mean absolute error of the results obtained using the DFA recommendation method to quantitatively determine whether the performance target has been achieved. This condition is directly related to the ultimate application goal of the DFA recommendation system, ensuring that the finally trained model set can effectively meet the preset accuracy requirements in actual chemical calculations. Attached Figure Description

[0020] Figure 1 Flowcharts illustrating training methods for density functional recommendation models according to several embodiments of this specification are shown.

[0021] Figure 2 A schematic diagram of the structure of an electronic device according to some embodiments of this specification is shown. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings.

[0024] The terms "first," "second," "third," etc., in the description, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0025] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes may be made to the function and arrangement of the described elements without departing from the scope of this specification. Various processes or components may be appropriately omitted, substituted, or added to the examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described with respect to some examples may be combined into other examples.

[0026] Figure 1 This specification illustrates several embodiments of training methods suitable for density functional recommendation models. The specific operational steps of this method 100 are described in detail below.

[0027] Box 101 is used to construct the seed training set for Method 100. This step takes a raw molecular library containing a large number of candidate molecules as input. First, cheminformatics methods such as k-medoids clustering are used to screen out a small number of representative molecules with structural diversity, forming an initial molecular subset. Subsequently, high-cost "gold standard" calculations and multiple density functional approximations are performed on each molecule in this subset to obtain the true error labels of its target attributes; simultaneously, a low-cost DFT calculation and density fitting are used to extract its atomized electronic density features. The final output is a structured initial training set, in which each sample contains the molecule's electronic density feature vector and its corresponding label vector composed of multiple DFA true absolute errors. The goal is to construct a seed training set with sufficient chemical diversity that can be used to initiate subsequent active learning processes with the lowest initial computational cost.

[0028] Specifically, the first step is to obtain a raw molecular library. This is a dataset containing a large amount of candidate molecular structure information. This library can be derived from public databases, such as the Cambridge Structural Database or CSD, or a virtual molecular library generated according to specific chemical rules, such as octahedral coordination or the inclusion of specific 3d transition metals. For example, this embodiment begins with a raw molecular library containing approximately 10,000 three-dimensional structures of transition metal complexes.

[0029] Box 101 requires software for performing high-precision quantum chemical calculations. For example, the "gold standard" calculation can use the DLPNO-CCSD(T) method in ORCA software, and density functional approximation calculations can use Psi4 or TeraChem software. A list containing N candidate DFAs, basis sets (such as def2-TZVP), and other calculation parameters must be predefined. For example, N=48.

[0030] The process is divided into two sub-stages: representative molecule screening and high-cost computational labeling.

[0031] Representative molecule screening involves selecting a small subset with the greatest structural diversity from the original molecular library using cheminformatics methods. This selection can be done manually based on experience or using clustering algorithms. For example, the k-medoids clustering algorithm can be used to cluster approximately 10,000 molecules in the library based on their physicochemical descriptors, such as metal center type, ligand type, and expected ligand field strength. Then, a representative molecule is selected from the center of each cluster.

[0032] For example, we set an initial selection of M0 = 20 molecules to form an initial molecular subset. These 20 molecules will cover different metals, such as Fe and Co, oxidation states, and typical ligands, such as... , Cl ensures the breadth of the chemical space.

[0033] Then, high-cost computational labeling is performed. For each molecule in the initial subset of molecules selected, the following two types of calculations are performed in sequence to obtain its true error label.

[0034] a. Gold Standard Calculation: The target property of the molecule is calculated using the high-precision DLPNO-CCSD(T) method. In this embodiment, the target property is the vertical spin splitting energy.

[0035] b. Multiple DFA calculations: Using 48 predefined density functional approximations, the same target property of the same molecule, namely the vertical spin splitting energy mentioned above, is calculated separately.

[0036] Subsequently, for each molecule and each DFA, its true absolute error is calculated:

[0037] True absolute error = |DFA calculated value - Gold standard calculated value|.

[0038] Simultaneously, a low-cost feature calculation is performed for each molecule: a general-purpose, computationally inexpensive DFA, such as B3LYP, is used. Single-point calculations are performed on the optimized structure, and atomized electron density features invariant to rotation, translation, and substitution, i.e., the power spectrum, are extracted from its electron density using density fitting techniques and used as machine learning input features for the molecule.

[0039] After the operations described above, box 101 outputs a structured initial training set. This training set contains 20 data samples, each corresponding to an initial molecule, and includes a feature vector and a label vector. The feature vector represents the atomized electron density feature of the molecule. The label vector is a real vector of length N (48), where the i-th element is the true absolute error of the molecule corresponding to the i-th DFA.

[0040] A specific molecule in the initial molecular subset For example, in box 101, we calculate the reference value of its vertical spin splitting energy using DLPNO-CCSD(T) as -ΔE_ref kcal / mol. We calculate using 48 different DFAs, obtaining 48 values, such as the calculated value for the PBE0 functional as -ΔE_PBE0 kcal / mol. Therefore, for the PBE0 functional, its true absolute error label is |-ΔE_PBE0 - (-ΔE_ref)| kcal / mol.

[0041] Meanwhile, the electron density eigenvector F_Fe is extracted from B3LYP calculations.

[0042] Ultimately, the molecule is stored as a single record in the initial training set: {feature: F_Fe, label: [error_DFA1, error_DFA2, ..., error_DFA48]}.

[0043] Box 102 is used to implement the training initial error prediction model set of method 100.

[0044] In box 102, using the initial training set constructed in box 101, an initial error prediction model group corresponding to multiple density functional approximations (DFAs) is trained through supervised learning, providing a basic discriminative ability for subsequent active learning screening.

[0045] Specifically, the initial training set output from box 101 is used as input. Each sample in this training set contains an atomized electron density feature vector of a molecule and a corresponding true absolute error label vector. A set of neural networks is used to train supervised learning on this dataset. For example, for the 48 candidate DFAs preset in this embodiment, a neural network with the same structure is initialized, such as using a modified Behler-Parrinello architecture. Each network's task is independent regression prediction: taking the electron density feature vector of a molecule as input, it learns to map the true absolute error value of that molecule to the specific DFA that the network is responsible for. For example, for the model responsible for predicting the PBE0 functional error, the label in its training data is the error value corresponding to PBE0 in each sample label vector.

[0046] During training, the mean squared error loss function and the Adam optimizer are used to minimize the gap between the network's predicted values ​​and the true labels.

[0047] The Mean Squared Error (MSE) loss function calculates the squared difference between the predicted and true values. This means that when the model's DFA error prediction for a molecule deviates significantly, the loss value for that sample increases quadratically, thus attracting greater attention during training. The Adam optimizer, on the other hand, combines momentum and adaptive learning rate. It maintains an independent learning rate for each parameter and dynamically adjusts it based on the mean of historical gradients and the uncentered variance. This mechanism effectively accelerates convergence in the early stages of training, significantly reducing the complexity and time cost of model parameter tuning. Therefore, MSE ensures the model takes large errors seriously, while Adam ensures the model can efficiently and stably find better solutions.

[0048] With the aforementioned molecules For example, its electron density feature F_Fe is simultaneously input into 48 networks, but each network receives a different training objective: the first network learns the error corresponding to DFA1 in the label vector, the second network learns the error corresponding to DFA2, and so on. Through this process, each network gradually learns to identify chemical patterns from the electron density feature that may cause computational errors in its specific DFA.

[0049] The final output of box 102 is an initial error prediction model set, which contains 48 neural network models that have undergone preliminary training. Although the prediction accuracy and generalization ability of the models trained on the small initial training set are still limited, they already have basic error trend prediction capabilities, which are sufficient to drive the effective screening of high uncertainty molecules in the next box.

[0050] In box 103, the initial error prediction model group trained in box 102 is used to evaluate a large number of unlabeled molecules in the original molecular library that have not yet been labeled with high cost. The aim is to screen out a group of molecules with the greatest prediction discrepancy and the highest uncertainty as the optimal candidates for the next round of high cost computation labeling, thereby maximizing the data acquisition efficiency of active learning.

[0051] Specifically, the input for this step mainly includes two parts: first, all unlabeled molecules in the original molecular library that were not selected into the initial training set in box 101, such as the remaining approximately 9,980 transition metal complex structures in the example above; second, the initial error prediction model set output in box 102, namely the 48 pre-trained DFA error prediction neural networks.

[0052] The core of the process is uncertainty assessment and screening. First, for each unlabeled molecule, a low-cost feature extraction calculation, similar to that in box 101, needs to be performed. For example, B3LYP functional calculations can be used to extract its atomized electron density features.

[0053] Then, the molecule's feature vector is simultaneously input into 48 initial error prediction models, and each model outputs a prediction absolute error value for that molecule, corresponding to its specific DFA.

[0054] At this point, for each unlabeled molecule, a vector consisting of 48 predicted values ​​has been obtained. Next, a comprehensive uncertainty measure is calculated based on this set of predicted values.

[0055] In active learning, the goal is to obtain the best-performing model with the fewest labeled samples. Therefore, priority is given to labeling samples that can best train the model and provide the most new knowledge. Samples whose predictions the model is less certain usually mean that their features are at the boundary or blind spot of the model's current knowledge domain—precisely the information the model most needs to learn. Using uncertainty metrics maximizes information acquisition efficiency. Concentrating expensive computational resources on labeling the most uncertain molecules allows each unit of computation to bring the greatest performance improvement. This avoids wasting resources labeling molecules that the model can already accurately predict or that are highly similar to existing data.

[0056] For example, the statistical variance of these 48 predicted values ​​can be calculated. A larger variance indicates greater discrepancy in the predictions of different models regarding the molecule, meaning the model set is more uncertain about its understanding of the molecule, and the potentially higher the amount of information it contains. Variance is the most direct statistic for measuring the dispersion of a set of data.

[0057] Furthermore, variance calculation is a fundamental mathematical operation with extremely low computational cost, making it virtually effortless even when evaluating a large number of molecules. Moreover, the physical meaning of variance calculation is very clear. Large variance directly indicates significant discrepancies in the error assessments of different DFA expert models. This discrepancy may stem from a molecule possessing a rare electronic structure feature, leading to substantial differences in the performance of different theoretical approximations—a highly valuable learning opportunity. Furthermore, variance calculation utilizes only the model's standard output, requiring no additional confidence information from within the model, thus ensuring compatibility with any regression model architecture and strong universality.

[0058] For example, for an unlabeled cobalt complex molecule The model set may have very low error predictions for some of its DFAs (e.g., 1.2 kcal / mol) but very high predictions for others (e.g., 15.5 kcal / mol), resulting in a large prediction variance. This suggests that the chemical properties of the molecule may fall within the boundaries of the current model's knowledge domain.

[0059] The system calculates the predicted variance for all unlabeled molecules and sorts them in descending order based on the variance values. Then, according to a preset batch size, it selects the top K molecules with the largest variances from the sorted list to form a candidate list of molecules with high uncertainty (e.g., K=50). The final output of this step is this candidate list.

[0060] In other embodiments, variance may not be used; instead, other calculation methods may be employed to obtain the aforementioned list of high-uncertainty molecular candidates. For example, in some embodiments, entropy, standard deviation, etc., may be used.

[0061] Box 104 is used to implement the high-cost computational labeling of high-uncertainty molecules in method 100.

[0062] In box 104, for each molecule in the high uncertainty molecule candidate list selected in box 103, high-precision "gold standard" quantum chemical calculations and a full set of candidate density functional approximation (DFA) calculations are performed to obtain its accurate true error label.

[0063] The input to box 104 is the list of high-uncertainty molecular candidates output from box 103. This list contains the K molecular structures evaluated by the model group as having the highest prediction uncertainty. Continuing the previous example, this list may contain, for example, Molecules with special ligand fields or electronic structures are selected. The core of the processing is targeted, high-cost computation, which operates exactly the same as the operation on the initial subset of molecules in box 101, but the target molecules become this carefully selected batch of highly uncertain molecules.

[0064] For example, in the list of high-uncertainty molecular candidates... molecular:

[0065] First, the reference value Eref of its vertical spin splitting energy was calculated using the DLPNO-CCSD(T) method. Then, 48 EDFAj values ​​were calculated using 48 different DFAs, such as PBE, B3LYP, and M06-2X, with j = 1, 2, 3…48.

[0066] Then, apply the above formula to each DFA. For example, if the calculated value of the PBE0 functional, EPBE0, is -45.3 kcal / mol, while the reference value, Eref, is -48.1 kcal / mol, then the true absolute error of the PBE0 functional for this molecule is |-45.3-(-48.1)| = 2.8 kcal / mol. The eigenvectors of this molecule do not need to be recalculated; the previously extracted electron density features can be used.

[0067] Therefore, the final output of box 104 is the new training data for these K, for example, 50, high-uncertainty molecules. Each molecule's data record contains a feature vector and a label vector. Similarly, the feature vector is its atomized electron density feature. The label vector is a real-valued vector of length 48, where the j-th element is the true absolute error (ErrorDFAj) of the j-th DFA on that molecule.

[0068] Box 105 is the implementation of method 100, which updates the training set and iteratively optimizes the model group.

[0069] In box 105, the high-quality new training data generated in box 104 is merged with the historical training set, and the existing error prediction model set is retrained and optimized using this expanded training set.

[0070] Box 105 is the model evolution stage in the active learning loop. Its purpose is to efficiently integrate newly acquired knowledge into the model, thereby systematically improving the overall prediction accuracy and generalization ability of the model group.

[0071] The input to box 105 includes the historical training set and the newly added training data.

[0072] The historical training set refers to the training set used after the previous iteration. When box 105 is executed for the first time, this is the initial training set output by box 101, which contains only 20 molecular samples; in subsequent iterations, it is the expanded training set output by the previous box 105. The new training data is the new data output by box 104, targeting K molecules with high uncertainty.

[0073] The model set to be optimized is the previous model set. During the initial execution, it's the initial error prediction model set output from box 102; subsequent runs use the optimized model set output from box 105. The core of the process is data integration and model retraining. First, the newly added training data is merged with the historical training set to form an expanded new training set. For example, if the initial training set has 20 samples and 50 new samples are added, the new training set contains 70 samples. Then, this expanded training set is used as a new source of supervisory signals to retrain the previous model set.

[0074] In different implementations, the specific training methods can vary. For example, it could be training from scratch, using the expanded complete training set to reinitialize and train a completely new set of 48 models. Alternatively, transfer learning / fine-tuning, which is more efficient, could be used. The parameters of the previous model set are used as initial values, and additional training rounds are conducted using the expanded training set. This allows the model to focus on learning new knowledge from the newly added samples, building upon existing knowledge, typically resulting in faster and more stable convergence.

[0075] The training algorithm is consistent with that in box 102, employing a supervised learning framework. It uses the electron density features of the molecule as input and the true absolute error of the corresponding DFA as the regression target, optimizing it using a neural network, mean squared error loss, and the Adam optimizer. For example, after learning new features... After obtaining precise error data for molecules, its internal parameters are adjusted to correct or enhance the ability to predict errors for molecules with similar chemical characteristics.

[0076] Box 106 is used to determine the termination of the iteration in method 100.

[0077] In box 106, the system evaluates whether the current active learning loop should continue. By determining whether the preset termination conditions are met, it decides whether to start a new round of "selection-labeling-training" loop or to end the process and output the final model. Therefore, box 106 is actually the decision point for controlling the final balance between training cost and model performance.

[0078] The input to box 106 is primarily the performance evaluation results of the next-generation model set output from box 105, along with optional records of computational resource consumption. Evaluation is typically performed on a fixed, independent validation set that has never participated in any training or selection.

[0079] The validation set needs to include a batch of molecules known as the "gold standard" and those calculated using DFA, for example, 152 molecules reserved from the original molecular library. The processing involves conditional judgment, mainly based on the following two types of preset conditions, and can terminate if either condition is met.

[0080] Condition one, namely, performance compliance. This involves determining whether the model set's performance has met the preset accuracy requirements. The core evaluation metric is the final calculation error using the DFA recommended by the model set. A DFA is recommended for each molecule in the validation set using the current model set, and then the mean absolute error between the result obtained using the recommended DFA and the "gold standard" result is calculated. It should be noted that the process of recommending DFAs for molecules using the model set is existing technology, and as indicated in the background section of this specification, the DFA with the smallest error is the recommended optimal DFA. If this error is lower than or equal to a preset threshold, for example, ≤ 2.1 kcal / mol, i.e., "chemical accuracy" is achieved, then the performance is considered to have met the requirements.

[0081] Condition two is the resource exhaustion condition. This involves determining whether the total number of high-cost computations or the total computation time performed to obtain training data has reached the preset budget limit. Specifically, the total number of computations using the "gold standard" plus multiple DFAs has reached its limit. For example, if the preset total number of molecules to be computed cannot exceed 200, and after multiple iterations the total number of labeled molecules has reached 190, then resource exhaustion is imminent.

[0082] If any termination condition is met, the process ends. The output instruction is termination, and the final set of error prediction models output by box 105 is delivered as the final tool that can be used for DFA recommendations.

[0083] If none of the termination conditions are met, the output instruction is to continue iteration. The process will jump back to box 103, using the currently amplified training set, the optimized model set, and the remaining unlabeled molecules in the molecular library as input, to start the next round of active learning loop.

[0084] For example, suppose after two rounds of iteration (initial 20 + 50 new + 50 new), the training set has 120 molecules. The current model's recommended MAE on the validation set is 2.3 kcal / mol, while the preset performance threshold is 2.1 kcal / mol, and the resource limit is 200 labeled molecules. Since the performance is not up to standard and resources are not exhausted, the decision is "continue iterating," and a third round of high-uncertainty molecule screening will be performed. If, in subsequent iterations, the model's recommended MAE drops to 2.0 kcal / mol, the decision is "termination," and the output model at this point is the final version.

[0085] Based on the method described above 100, this solution fundamentally reconstructs and optimizes the model training process through an intelligent iterative data acquisition strategy, achieving an order-of-magnitude reduction in training costs, thereby greatly enhancing the practicality and scalability of the recommendation technology.

[0086] Taking a training set containing 452 molecules as an example, existing technologies require extremely expensive "gold standard" quantum chemical calculations and a full suite of multi-DFA calculations for these 452 molecules. However, in the method 100 of this invention, only a small number of representative molecules (e.g., 20) and a small batch of molecules with the highest model prediction uncertainty (e.g., 50 per round) are subjected to the aforementioned high-cost calculations to obtain their precise labels. In this way, the system accurately allocates limited computational resources to the data points that contribute most to improving model performance.

[0087] Therefore, this approach can train a set of high-precision DFA error prediction models with performance comparable to those trained using all samples, at the cost of high-cost labeling of only a small subset of all molecules. This means that the total computational cost of the model training phase can be reduced to less than half that of the original approach. This order-of-magnitude cost reduction transforms the construction of high-performance DFA recommendation systems from a cutting-edge task relying on top-tier computing resources into an economically feasible process within a regular research budget, effectively solving the core bottleneck hindering the widespread adoption and application of this technology.

[0088] Figure 2 A block diagram of an electronic device 300 that can implement various embodiments of the present disclosure is shown. For example... Figure 2 As shown, device 300 includes a processor 301, which can perform various appropriate actions and processes based on computer program instructions loaded into random access memory (RAM) 303 according to computer program instructions stored in read-only memory (ROM) 302. RAM 303 may also store various programs and data required for the operation of device 300. The processor 301, ROM 302, and RAM 303 are interconnected via bus 304. Input / output (I / O) interface 305 is also connected to bus 304.

[0089] The various processes and procedures described above, such as method 100, can be executed by processor 301. For example, in some embodiments, method 100 may be implemented as a software program tangibly contained in a machine-readable medium. In some embodiments, part or all of the software program may be loaded and / or installed on device 300 via ROM 302. When the software program is loaded into RAM 303 and executed by processor 301, one or more actions of method 300 described above may be performed.

[0090] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0091] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. Furthermore, although operations are depicted in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the foregoing discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0092] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A training method suitable for density functional recommendation models, characterized in that, It includes the following steps: The original molecular library is obtained, and some representative molecules are selected as the initial molecular subset. For each molecule in the subset, a feature vector and a tag vector are calculated. The feature vector is the atomized electron density feature of the molecule, and the tag vector is a real number vector of length N, where N is the number of candidate DFAs. The real number vector contains the true absolute error of the molecule for each DFA. The molecular subset and the corresponding feature vector and label vector constitute the initial training set; Create an error prediction model group containing N error prediction models; use the initial training set to train the error prediction model group to obtain the trained initial model group; The feature vector of each molecule is calculated for the molecules in the original molecular library that were not selected into the initial molecular subset. This feature vector is then placed into the initial model group for error prediction to obtain the uncertainty metric for each molecule. Molecules with high uncertainty metrics are selected as the optimal candidate molecules for the next round. The label vector of the optimal candidate molecule is calculated, and the optimal candidate molecule is used as new training data to update the training set. The initial model group is then trained iteratively to obtain a new round of error prediction model group.

2. The training method for density functional recommendation models according to claim 1, characterized in that: In the process of selecting a representative subset of molecules as the initial molecular subset, a clustering algorithm is used to cluster the original molecular library based on the physicochemical descriptors of the molecules. Representative molecules are selected from the center of each cluster to form the initial molecular subset.

3. The training method for density functional recommendation models according to claim 1, characterized in that: The true absolute error is defined as: |DFA calculated value - gold standard calculated value|, where the DFA calculated value is the corresponding target attribute of the molecule calculated by the DFA, and the gold standard calculated value is the corresponding target attribute of the molecule calculated using the high-precision DLPNO-CCSD method.

4. The training method for density functional recommendation models according to claim 1, characterized in that: The creation of the error prediction model group specifically involves initializing a neural network for each of the N preset DFAs. Each neural network is configured to perform independent regression prediction, taking the electron density feature vector of the molecule as input, and learning to map the true absolute error value of the molecule to the specific DFA that the neural network is responsible for.

5. The training method for density functional recommendation models according to claim 1, characterized in that: During the training of the error prediction model group, the mean squared error loss function and the Adam optimizer are used to minimize the gap between the neural network predictions and the true labels.

6. The training method for density functional recommendation models according to claim 1, characterized in that: The uncertainty measure for each molecule is obtained by simultaneously inputting the molecule's feature vector into N error prediction models. Each error prediction model outputs a prediction absolute error value for that molecule, corresponding to its specific DFA. The uncertainty measure is obtained based on these N prediction absolute error values.

7. The training method for density functional recommendation models according to claim 6, characterized in that: The uncertainty measure of this molecule is achieved by calculating the statistical variance of these N absolute error values ​​of prediction; the larger the variance, the higher the uncertainty measure.

8. The training method for density functional recommendation models according to any one of claims 1-7, characterized in that: After obtaining a new set of error prediction models, it is necessary to determine whether to iterate the error prediction model set again. If the performance target is met or the resource exhaustion condition is met, the error prediction model set will not be iterated again.

9. The training method for density functional recommendation models according to claim 8, characterized in that: Create an independent validation set, use the current error prediction model set to recommend DFA to the molecules in the validation set, and calculate the mean absolute error between the result obtained by using the DFA and the gold standard result. If it is lower than the preset threshold, it is determined that the performance meets the standard.

10. An electronic device, comprising a processor and a memory; the processor being connected to the memory; The memory is used to store executable program code; the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to perform the method as described in any one of claims 1 to 9.

Images