Method and apparatus for characterizing chemical substance, measuring physicochemical properties and generating control data for synthesizing chemical substance

Through data-driven compression model and neural network technology, multimodal representation of chemical substances is generated, which solves the problem of frequent measurements and waste in the synthesis of new polymers in the chemical industry, and realizes efficient chemical substance representation and resource optimization .

CN119032401APending Publication Date: 2024-11-26BASF SE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202380033812.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-12-27
Filing Date
2023-04-14
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In the chemical industry, due to the complexity of characteristics in the synthesis of new polymers, the success rate is low, wasteful, time-consuming and expensive. The existing technology is difficult to effectively reduce the number of polymer measurements and cannot provide diversified chemical representations method.

Method used

Using a data-driven compression model, multimodal data is encoded and decoded through neural networks to generate a predetermined multimodal representation, characterizing the physical and chemical characteristics, composition and identifiers of chemical substances, reducing the dimensions of the measured data, and improving processing efficiency.

Benefits of technology

The efficient digital representation of chemical substances is achieved, reducing the number of polymer measurements, improving the efficiency and resource utilization of the chemical process, and reducing production costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119032401A_ABST
    Figure CN119032401A_ABST
Patent Text Reader

Abstract

A method for characterizing a chemical substance in a predetermined multi-modal representation, the predetermined multi-modal representation having a predetermined plurality of modalities, the method comprising: receiving (S12) multi-modal substance data (2, 12) comprising a first set of modalities of the chemical substance; encoding (S13) said multi-modal data (2, 12) using a data-driven model (3, 10) of the chemical substance to generate encoded substance data; and generating multi-modal substance data comprising a second set of modalities of the chemical substance by decoding the encoded substance data using the data-driven model (3, 10), where the multi-modal substance data is indicative of a physicochemical property of the chemical substance, a composition of the chemical substance and / or an identifier of the chemical substance. The data-driven model (3, 10) is implemented to map input data (2, 12) to encoded output data (6), the input data being a multi-modal representation (2, 12) of the chemical substance. The encoded output data is a potential spatial representation (6) of the input data (2, 12). Further, the first set of modalities is different from the second set of modalities, and the first set of modalities and the second set of modalities are included in the predetermined plurality of modalities. At least one modality in the second group is not included in the first group.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The present disclosure relates to a method for generating a digital representation of a chemical substance. This may involve aspects of training a neural network to represent the chemical substance. The present disclosure further relates to a computer program product and to a database search device for identifying a chemical substance. Applications of the digital representation involve generating measurement data associated with the chemical substance and control data associated with synthesis specifications of the chemical substance. The disclosed methods and aspects relate to characterizing the chemical substance in terms of physicochemical properties, composition of the chemical substance, and / or identifiers.

[0002] Chemical substances, such as polymers, come in many shapes, sizes, and compositions. Small molecules are usually represented by their structure (chemical composition). Other basic chemicals can sometimes be represented by their formula and detailed description of their manufacturing process. However, the characteristics of chemical substances, such as polymers, are often too complex to be represented by their formula or their structure. It is desirable to provide more diverse ways to represent chemical substances.

[0003] In particular, in the chemical industry, chemicals such as new polymers are increasingly being customized to customer requirements. This requires synthesizing new polymers and then performing measurements on various characteristics. This is very expensive, and synthesizing polymers often generates unnecessary waste because the success rate of synthesizing polymers that meet customer requirements is low. In addition, performing measurements is time-consuming and expensive. Therefore, there is a need to reduce the number of measurements required to fully analyze a polymer. Digital representations of chemical substances can be a means for reducing the effort made in determining or measuring the physical or chemical properties of chemical substances, generating specifications for synthesized chemical substances, and / or identifying chemical substances for specific purposes.

[0004] It is an object of the present invention to provide methods and apparatus for generating digital representations of chemical substances. Further objects include improving the use and application of digital representations in the context of chemical substance synthesis and substance characterization.

[0005] The aforementioned objects are met by a method and an apparatus according to the independent claims.

[0006] According to one aspect, a method for characterizing a chemical substance in a predetermined multimodal representation is provided. The predetermined multimodal representation has a predetermined plurality of modalities, and the method comprises the following steps:

[0007] receiving multimodal material data including a first set of modalities of the chemical substance;

[0008] encoding the multimodal data using a data-driven (particularly compression) model of the chemical substance to generate encoded substance data; and

[0009] generating a predetermined multimodal representation comprising a second set of modalities of the chemical substance by decoding the encoded substance data using a data-driven compression model, wherein, preferably, the multimodal substance data indicates a physicochemical property of the chemical substance, a composition of the chemical substance and / or an identifier of the chemical substance;

[0010] wherein the data-driven compression model is implemented to map input data to encoded output data, the input data being a multimodal representation of the chemical substance, and the encoded output data being a latent space representation of the input data;

[0011] Wherein, preferably, the first group of modalities is different from the second group of modalities, and wherein the first group of modalities and the second group of modalities are included in the predetermined plurality of modalities, wherein, in particular, at least one modality in the second group is not included in the first group.

[0012] Modality can be considered as a data set indicating a specific type of information (particularly physicochemical properties) associated with a chemical substance. Many different types of data are included in a corresponding multimodal data set as multimodal data. Modality can be represented by a dedicated data structure or a data type with a specific number of dimensions. A scalar or multidimensional vector or tensor can be envisioned as a data type. For example, melting temperature can be considered as a physicochemical property as a modality of a chemical substance. Another modality can be associated with a SMILES (Simplified Molecular Linear Input Specification) representation of a substance. Identifiers such as names (strings) or chemical formulas can be used as modal data. In an embodiment, each modality is represented by a specific data structure so that computer-implemented processing is performed on the modal data. The data structure can contain a set of parameters that characterize the substance.

[0013] In one aspect, the method allows generating output multimodal material data from input material data that is "incomplete" with respect to a predetermined multimodal representation. For example, the input data includes a first modality and other modalities, but does not include a second modality. Due to the data-driven compression model, the generated output material data may include the second modality. The method may be understood as providing measurement data from input data.

[0014] In an embodiment, the first set of multiple modalities is a subset of the second set of multiple modalities.

[0015] In various aspects, the method allows performing a process using a latent space representation of input data. The latent space representation can be thought of as a vector in a latent space. The process within the latent space can involve searching for vectors, comparing vectors associated with different chemicals in the latent space, defining regions in the latent space, and calculating similarity measures or scores.

[0016] In an embodiment, the data-driven compression model includes at least one trained neural network, which is implemented to receive multimodal input data including the predetermined multiple modalities, encode the input data into a latent space representation of the input data, and decode the encoded input data into multimodal output data including the predetermined multiple modalities.

[0017] In an embodiment, the neural network is trained based on training data, the training data comprising multimodal training data, the multimodal training data comprising the predetermined multiple modalities. For example, the neural network is trained using all available modalities in the predetermined multiple modalities.

[0018] In an embodiment, the neural network comprises a plurality of individual encoders, wherein each individual encoder is assigned to one of the predetermined plurality of modalities, wherein each individual encoder is implemented to bring input data from the modality to which the individual encoder is assigned into the same dimensionality of the latent space.

[0019] In an embodiment, the neural network comprises a plurality of individual decoders, wherein each individual decoder is assigned to one modality of the predetermined plurality of modalities, wherein each individual decoder is implemented to decode a latent space representation of the encoded input data into modal data in the generated multimodal material data, the modality being the modal data of the modality to which the individual decoder is assigned.

[0020] In an embodiment, characterizing comprises measuring a physicochemical property of a chemical substance, the substance data comprises sensor data, the measuring comprising:

[0021] receiving sensor data indicative of a first measurable physicochemical property of the chemical substance, the sensor data being associated with at least one modality included in the first group;

[0022] encoding the sensor data using a data-driven compression model of the chemical substance to generate encoded sensor data; and

[0023] Measurement data indicative of a second measurable physicochemical property of the chemical substance is generated by decoding the encoded sensor data using the data-driven compression model, the measurement data being associated with at least one modality included in the second group.

[0024] In an embodiment, the chemical substance is a polymer.

[0025] The method can be viewed as a method for measuring observable quantities of chemical substances, where the observable quantities are not directly available but can be derived or reproduced by a data-driven model. The data-driven model inherently acquires knowledge about the expected output measurement data by being trained in advance using "complete" material training data.

[0026] In an embodiment, at least one item of the group consisting of: the generated measurement data, the recipe data indicating the chemical substance, the identification data indicating the chemical substance is output.

[0027] In an embodiment, the method comprises the following steps:

[0028] For a plurality of sample chemicals, using the data-driven compression model to generate a latent space representation of multimodal sample material data associated with the sample chemical substances, the multimodal sample material data having the predetermined plurality of modalities; and / or

[0029] The generated latent space representation of the multimodal sample material data associated with the sample chemical substances is stored in a database.

[0030] In an embodiment, the method comprises the following steps:

[0031] receiving search input data indicating predetermined physicochemical properties of a chemical substance to be searched, wherein the search input data is provided as a multimodal representation of the substance to be searched included in the first group;

[0032] encoding received search input data to generate a latent space representation of the search input data; and

[0033] The generated latent space representation of the search input data is compared with the latent space representation of the sample chemical substances to obtain a comparison result.

[0034] In an embodiment, the method comprises the following steps:

[0035] In response to the comparison, at least one sample chemical substance is selected.

[0036] In an embodiment, wherein comparing comprises:

[0037] calculating a similarity score of the latent space representation of the search input data relative to the latent space representation of the sample chemical; and / or

[0038] A similarity range within the latent space relative to the latent space representation of the search input data is determined.

[0039] In an embodiment, at least one modality in the first group and / or the second group includes synthesis specifications of the chemical substance, and / or control data indicative of synthesis specifications of the chemical substance.

[0040] In an embodiment, characterizing comprises generating control data indicative of synthesis specifications of the chemical substance, in particular a polymer, the generating comprising:

[0041] providing a first synthesis specification of a reference chemical substance as at least one modality in the first group;

[0042] encoding the first synthesis specification into a digital representation of the reference chemical substance using the data driven compression model;

[0043] providing a database including a plurality of historical digital representations of historical chemical substances;

[0044] determining a similarity score of the historical digital representations relative to the digital representation of the reference chemical;

[0045] Based on the similarity score, selecting at least one historical representation, and decoding to generate a synthesis specification associated with the at least one selected historical representation; and

[0046] Control data indicating the generated synthesis specification is generated.

[0047] In an embodiment, these historical chemistries are sample chemistries.

[0048] In an embodiment, characterization comprises generating control data indicative of synthesis specifications of a chemical substance, in particular a polymer, the generating comprising:

[0049] receiving sensor data indicative of a measurable physicochemical property of the chemical substance;

[0050] encoding the sensor data using a data-driven compression model of the chemical substance to generate encoded sensor data; and

[0051] Control data indicating synthesis specifications of the chemical substance is generated by decoding the encoded sensor data using the data driven compression model.

[0052] In an alternative embodiment, the first set of modalities is identical to the second set of modalities. Thus, the input and output modalities may be identical. However, the output data may refer to a different polymer or substance, or material property data, than the input search.

[0053] In an embodiment, using the data driven model includes a process for generating a compressed digital representation of a chemical substance, particularly a polymer, comprising:

[0054] receiving input data that is a multimodal representation of a physicochemical property of the chemical substance and is indicative of a measurable physicochemical property of the chemical substance;

[0055] encoding the input data using a data-driven compression model of the chemical substance to generate encoded substance data based on the received input data; and

[0056] Chemical substance data indicating measurable physicochemical properties of the chemical substance is generated by decoding the encoded substance data using the data-driven compression model.

[0057] In an embodiment, the process is implemented as a method according to the second aspect disclosed below.

[0058] According to a second aspect, there is provided a method for generating a representation of a chemical substance, in particular a polymer, the method comprising:

[0059] receiving input data that is a multimodal representation of a physicochemical property of the chemical substance and is indicative of a measurable physicochemical property of the chemical substance;

[0060] encoding the input data using a data-driven compression model of the chemical substance to generate encoded substance data based on the received input data; and

[0061] Chemical substance data indicating measurable physicochemical properties of the chemical substance is generated by decoding the encoded substance data using the data-driven compression model.

[0062] The method of the second aspect is particularly the process involved when using a data-driven model.

[0063] The method of the second aspect allows efficient processing of data relating to properties of chemical substances. Using multimodal data in an encoded or compressed form by using a data-driven model can increase processing speed and reduce the amount of resources required in terms of energy, materials and / or computing power.

[0064] In an embodiment, a data-driven compression model maps multiple input modalities into a latent space vector. Multimodal data including the modalities requires a first amount of data, and a latent space representation of the multimodal data requires a second amount of data. Preferably, the second amount of data is less than the first amount of data. Thus, using a data-driven model reduces the amount of data to be processed in the latent space without substantially losing information about the substance associated with the input multimodal data.

[0065] Multimodal data related to chemical substances include, for example, various data representing different aspects of the chemical substance. It can be envisioned that the first mode or mode is a physical observable, such as melting temperature, hardness, acidity, etc., and the second mode or mode is a structural aspect, such as the chirality of an enantiomer. Spectral aspects can also be considered as modalities.

[0066] In an embodiment, the data driven compression model is implemented as a trained neural network. For example, the encoding step may include providing the input data as a multimodal representation as an input to the trained neural network.

[0067] According to one aspect, a method for providing a data-driven compression model is disclosed. The method comprises:

[0068] providing multimodal input data as input to a neural network to be trained, the multimodal input data providing a multimodal representation of the chemical in a multimodal initial space, the neural network comprising a multimodal variational autoencoder, the multimodal variational autoencoder comprising a multimodal encoder and a multimodal decoder, wherein the multimodal encoder comprises multimodal encoder layers having multimodal encoder weights defining how the multimodal encoder layers transform data, and wherein the multimodal decoder comprises multimodal decoder layers having multimodal decoder weights defining how the multimodal decoder layers transform data;

[0069] Using the multimodal encoder layers, adjusting the dimensionality of the multimodal input data to obtain multimodal latent data in a latent space;

[0070] Decoding the multimodal latent data using the multimodal decoder layers to obtain multimodal reconstructed data in the multimodal initial space;

[0071] Calculating a loss function of the multimodal variational autoencoder under a current set of multimodal encoder and decoder weights based on the multimodal input data and the multimodal reconstruction data; and

[0072] A neural network with a current set of multimodal encoder and decoder weights is provided as a data-driven compression model.

[0073] According to one aspect, another method for generating a representation of a chemical substance, particularly a polymer, in a latent space representation is provided. The method may involve training a neural network and comprises:

[0074] providing multimodal input data as input to a neural network to be trained, the multimodal input data providing a multimodal representation of the chemical in a multimodal initial space, the neural network comprising a multimodal variational autoencoder, the multimodal variational autoencoder comprising a multimodal encoder and a multimodal decoder, wherein the multimodal encoder comprises multimodal encoder layers having multimodal encoder weights defining how the multimodal encoder layers transform data, and wherein the multimodal decoder comprises multimodal decoder layers having multimodal decoder weights defining how the multimodal decoder layers transform data;

[0075] Using the multimodal encoder layers, adjusting the dimensionality of the multimodal input data to obtain multimodal latent data in a latent space;

[0076] decoding the multimodal latent data using the multimodal decoder layers to obtain multimodal reconstructed data in the multimodal initial space; and

[0077] A loss function of the multimodal variational autoencoder under a current set of multimodal encoder and decoder weights is calculated based on the multimodal input data and the multimodal reconstruction data.

[0078] In an embodiment, the method comprises the step of outputting configuration data indicative of the trained neural network and / or a latent space representation of the chemical substance as a digital representation of the chemical substance.

[0079] The digital representation may further include an identifier of the chemical substance, a plurality of modalities associated with the chemical substance in the form of measurable data of the chemical substance, and / or specification data for synthesizing the chemical substance in the form of recipe data.

[0080] Using a multimodal variational autoencoder for multimodal input data allows processing data from multiple modalities at once, and groups modalities for the same chemical together to represent the chemical in a single latent space representation. Thus, the neural network is trained to provide a reliable representation of the chemical in the latent space.

[0081] The generated digital representation requires a relatively small amount of data relative to a complete description including all measurable properties of the chemical substance, and thus reduces the amount of computational and memory resources required to process the data indicative of the corresponding chemical substance. Thus, the trained neural network can be considered a data-driven compression model.

[0082] The combination of a multimodal decoder and an encoder can be viewed as an embodiment of a data-driven compression model suitable for generating digital representations of chemical substances. In particular, an autoencoder as disclosed herein can be viewed as an embodiment of a data-driven compression model.

[0083] A chemical substance can be a form of matter with a constant chemical composition and characteristics. Examples of chemical substances include polymers and molecules. A polymer can be made up of many monomers joined together.

[0084] A neural network trained according to the method of the first aspect allows for providing a representation of a chemical substance in a latent space representation. A latent space may be an abstract multidimensional space that maps what the neural network has learned from its training data. A latent space representation may be a mathematical representation of the training data with an adjusted (usually reduced) dimensionality. "Adjusting" the dimensionality may mean increasing, decreasing the dimensionality or maintaining the dimensionality the same, in particular to reach a desired (predetermined) dimensionality.

[0085] As used herein, a modality is information related to a chemical substance from a specific source and / or sensor. Different modalities may be information related to a chemical substance from different sources and / or sensors. Different modalities may be images of the chemical substance obtained by a camera, spectral images of the chemical substance, formulations of the chemical substance, simulation data of the chemical substance, test data from tests on the chemical substance, etc. Information (data) from multiple modalities may be represented as multimodal input data.

[0086] In particular, the dimensionality of a multimodal input is defined by the nature of the measurements. For example, in spectroscopy, the response of a chemical is measured at different wavelengths, and the range of wavelengths is fixed to the region where the response of the chemical is expected to be seen.

[0087] The multimodal initial space may correspond to a space in which multimodal input data is provided to the neural network. The multimodal initial space may be defined by its dimensionality. The multimodal initial space may be a space in which data from different sources and / or sensors are provided directly from these sources and / or sensors, or in which the data has undergone some modification (e.g., conditioning and / or preprocessing).

[0088] A multimodal variational autoencoder (multimodal VAE) may be a variational autoencoder that combines and / or considers data from different modalities. The structure of a multimodal variational autoencoder may be similar to that of a standard variational autoencoder.

[0089] The multimodal encoder may be configured to receive multimodal input data as input and adapt its dimensionality to obtain multimodal latent data in a latent space. In particular, the dimensionality of the multimodal latent data is smaller than the dimensionality of the initial multimodal input data.

[0090] The multimodal decoder can be used to decode back the data encoded by the multimodal encoder. In other words, the multimodal decoder receives multimodal latent data as input and outputs multimodal reconstructed data in a multimodal initial space (ie, in the same space as the initial multimodal input data).

[0091] During training, a loss function of the multimodal variational autoencoder is calculated. The loss function may indicate the performance of the multimodal variational autoencoder. The loss function may correspond to the difference between the multimodal input data and the multimodal reconstruction data. The loss function may be determined by a mixture of experts or by an expert multiplication technique. In particular, the smaller the loss function, the better the multimodal variational autoencoder.

[0092] The current set of multimodal encoder and decoder weights relates to a set of multimodal encoder and decoder weights for a current run (iteration) of neural network training.

[0093] The trained neural network can be used to receive training data or other types of multimodal data related to known chemicals as input, and to provide a latent space representation thereof as output.

[0094] The applications of having a latent space representation of chemical substances provided by a trained neural network are further explained below. Examples include providing a search engine to search for similarities between chemical substances in the latent space. Other examples include the redesign of chemical substances and / or the design of new chemical substances.

[0095] Further, the latent space representation of chemical substances can be used, for example, in automated production processes for producing chemical substances. In fact, automated production machines (robots) usually require very rich and compact information about the chemical substances to be produced, which the latent space representation can provide.

[0096] According to a further embodiment, the method further comprises:

[0097] The multimodal encoder weights and / or the multimodal decoder weights are updated based on the loss function.

[0098] The results of the multimodal variational autoencoder, i.e., multimodal latent data and multimodal reconstructed data, can be changed by modifying the multimodal encoder weights and the multimodal decoder weights.

[0099] Thus, updating the multimodal encoder weights and / or the multimodal decoder weights based on the loss function allows modifying the result (output) of the multimodal variational autoencoder in order to reduce the loss function and / or improve the neural network during its training, in particular to provide a latent space representation that is as accurate and / or convenient as possible for the input data related to the chemical substances.

[0100] According to a further embodiment, the method further comprises:

[0101] The steps of providing multimodal input data, adjusting the dimensionality, decoding the multimodal latent data, calculating a loss function and / or updating the weights to reduce the loss function are repeated.

[0102] By repeating these steps, the loss function can be reduced more and more until a satisfactory and reliable neural network is obtained. After repeating (running) the steps of providing multimodal input data, adjusting the dimension, decoding multimodal latent data, calculating the loss function and / or updating the weights to reduce the loss function, etc. for a predetermined number of times, once the loss function is below a predetermined threshold, the training of the neural network can be terminated.

[0103] According to a further embodiment, the neural network further comprises a separate variational autoencoder assigned to each modality, the separate variational autoencoder comprising a separate encoder and a separate decoder, wherein the separate encoder comprises separate encoder layers having separate encoder weights defining how the separate encoder layers transform data, and wherein the separate decoder comprises separate decoder layers having separate decoder weights defining how the separate decoder layers transform data, the method further comprising:

[0104] Each individual variational autoencoder is trained as follows:

[0105] Inputting into the separate variational autoencoder separate input data representing only the assigned modality in the separate initial space;

[0106] Using the single encoder layer, adjusting the dimension of the single input data to obtain single latent data having a predetermined dimension;

[0107] decoding the single latent data using the single decoder layer to obtain single reconstructed data in the single initial space;

[0108] comparing the separate input data with the separate reconstructed data to obtain a comparison result;

[0109] updating the individual encoder weights and / or the individual decoder weights based on the comparison result; and

[0110] Repeating the steps of inputting separate input data, adjusting the dimension of the separate input data, decoding the separate latent data, comparing the data, and updating the separate encoder weight and / or the separate decoder weight to reduce the comparison result;

[0111] And the method further comprises:

[0112] The individual latent data from multiple individual variational autoencoders are used as multimodal input data of the multimodal variational autoencoder.

[0113] Before data from different modalities are input to a multimodal variational self-encoder, a separate variational self-encoder can be used to preprocess the data. For example, a separate variational self-encoder performs preprocessing or adjustment on separate data representing separate modalities, respectively. A separate variational self-encoder can bring data from all modalities into the same format and / or the same dimension (corresponding to a "predetermined dimension") of separate potential data. This can facilitate processing by a multimodal variational self-encoder that receives separate potential data and / or information about separate variational self-encoders (e.g., hyperparameters determined during training of separate variational self-encoders) as multimodal input data.

[0114] Each individual variational autoencoder (individual VAE) can be a variational autoencoder that only considers data from a single modality (the assigned modality). In this way, data from each modality is processed separately by a single individual variational autoencoder. The structure of the individual variational autoencoders can be similar to that of the standard variational autoencoder.

[0115] The separate encoders may be configured to receive as input separate input data of the assigned modality and adjust (increase, decrease, or maintain) the dimensionality thereof to obtain separate latent data in separate latent spaces.

[0116] A separate decoder can be used to decode back data encoded by a separate encoder. In other words, a separate decoder receives separate latent data as input and outputs separate reconstructed data in a separate original space (ie, in the same space as the original separate input data).

[0117] During training, a separate loss function is calculated for a separate variational self-encoder. The separate loss function may indicate the performance of the separate variational self-encoder. The separate loss function may correspond to the difference between the separate input data and the separate reconstructed data, and is represented as a comparison result. In particular, the smaller the comparison result, the better the separate variational self-encoder.

[0118] Thus, updating individual encoder weights and / or individual decoder weights based on the loss function allows the results (outputs) of the individual variational autoencoders to be modified in order to reduce the loss function and / or improve the neural network under training.

[0119] According to a further embodiment, the method further comprises:

[0120] Repeating the training of each individual variational autoencoder for different predetermined dimensions of the individual latent data; and

[0121] The predetermined dimension with the smallest loss function is selected as the predetermined dimension of the final trained neural network.

[0122] In particular, the loss function of the multimodal variational autoencoder is calculated for different multimodal input data sets corresponding to separate latent data of different predetermined dimensions obtained from separate variational autoencoders. By changing the predetermined dimensionality, the loss function can be reduced. The final trained neural network is the trained neural network. Preferably, the final trained neural network has hyperparameters and / or weights that minimize the loss function. The predetermined dimensionality can be considered as a hyperparameter.

[0123] According to a further embodiment, the loss function of the multimodal variational encoder is determined by a mixture of experts or by an expert multiplication technique.

[0124] The mixture of experts technique involves breaking down the predictive modeling task into subtasks, training an expert model on each subtask, developing a gating model that learns which expert to trust based on the input to be predicted, and combining the predictions. The multiplicative expert technique models a probability distribution by combining outputs from several simpler distributions.

[0125] According to a further embodiment, one of the modalities describing the chemical substance provides: a spectral representation, a rheological representation, a thermal representation, a chemical representation, a structural representation, a representation of solubility, a representation of dispersion, a representation of viscosity and / or a representation of surface tension of the chemical substance.

[0126] In particular, a modality can correspond to any analytical and / or characterization method of a chemical substance. These include: (i) all types of spectral data, such as Fourier transform infrared spectroscopy (FTIR), Raman spectroscopy, Raman microscopy, UV spectroscopy, nuclear magnetic resonance spectroscopy (NMR), mass spectrometry (MS), gas chromatography-mass spectrometry (GCMS), etc.; (ii) thermal analysis data, such as dynamic mechanical analysis (DMTA), differential scanning calorimetry (DSC), thermogravimetric analysis (TGA), dynamic mechanical analysis (DMA), etc.; (iii) structural characterization data, such as X-ray diffraction; and (iv) data from unique analytical techniques that measure the physical properties of polymers, such as viscometers, surface tension, etc.

[0127] FTIR, for example, always includes errors from domain-specific sources. Therefore, neural networks are trained, for example, to be able to handle such errors and to be able to process chemical data in a more general way.

[0128] According to a further embodiment, at least two modalities provided in the single or multimodal representation of the multimodal input data comprise data in different formats.

[0129] Data in different formats may include images, tables, text, etc. Data in different formats may also include data having different dimensions and / or data structured or presented in different ways.

[0130] According to a further embodiment, at least a part of the multimodal and / or individual input data is sequential data.

[0131] To process sequential data, the neural network may be a convolutional neural network.

[0132] According to a further embodiment, the individual input data and / or the multimodal input data comprise enhancement data.

[0133] Augmented data may specify data artificially generated, for example, by generating slightly modified copies of initial data. This allows increasing the amount of training data (individual and / or multimodal input data) and thus improving the training of the neural network.

[0134] According to a further embodiment, the step of adjusting the dimensionality of the separate input data using the separate encoder layers comprises increasing the dimensionality of the separate input data for at least one modality and reducing the dimensionality of the separate output data for at least another modality.

[0135] Individual encoder layers may increase or decrease the dimensionality of individual input data depending on the modality to which they are assigned, to obtain individual data of a predetermined dimensionality. For example, data representing some modalities may be scalars (optionally with error bars), and their dimensionality is increased to achieve a predetermined dimensionality.

[0136] According to another aspect, a method for measuring a physicochemical property of a chemical substance, in particular a desired, preset and / or predetermined physicochemical property of the chemical substance, comprises the following steps:

[0137] receiving sensor data indicative of a first measurable physicochemical property of a chemical substance;

[0138] encoding the sensor data using a data-driven compression model of the chemical substance to generate encoded sensor data; and

[0139] Measurement data indicative of a second measurable physicochemical property of the chemical substance is generated by decoding the encoded sensor data using the data driven compression model.

[0140] Preferably, the encoding and decoding steps are performed according to the first aspect of the present disclosure or the corresponding steps of its embodiments.

[0141] In an embodiment, the data-driven compression model is implemented to map input data to encoded output data according to a first aspect of a method for generating a representation of a chemical substance, wherein the input data is a multimodal representation of the physicochemical properties of the chemical substance, and the encoded output data is a latent space representation of the input data.

[0142] Determining the properties of chemical substances in digital latent space representations can avoid the need to measure and experiment with samples of the corresponding substances in other ways. Further, the data indicating physicochemical properties generated using the presented data-driven methods and compression models can be used as training data for artificial intelligence purposes. Therefore, one aspect of the present disclosure is also the use of the generated measurement data as training data. Although sensor data can be obtained through hardware measurements, the generated measurement data can be considered as synthetic measurement data.

[0143] In an embodiment, the sensor data provided or received relates to a first type of physical measurement and the generated measurement data relates to a second type of physical measurement. For example, it is envisaged that X-ray diffraction data is provided as received sensor data and near infrared spectroscopy as generated measurement data. The digital representation can partially dispense with energy and resource-intensive technical processes for measuring material samples.

[0144] According to yet another aspect, a measuring device for measuring a physicochemical property of a chemical substance (particularly a desired, preset and / or predetermined physicochemical property of the chemical substance) comprises:

[0145] an interface device implemented to receive sensor data indicative of a first measurable physicochemical property of the chemical substance;

[0146] an encoder device implemented to encode received sensor data and generate and output encoded sensor data; and

[0147] A decoder device is implemented to generate measurement data indicative of a second measurable physicochemical property of the chemical substance and to decode the encoded sensor data.

[0148] The encoder device is preferably implemented to map input data to encoded output data according to a first aspect of a method for generating a representation of a chemical substance, wherein the input data is a multimodal representation of the physicochemical properties of the chemical substance in a multimodal initial space, and the encoded output data is a latent space representation of the input data.

[0149] The decoder device is preferably implemented to map input data to decoded output data according to the first aspect of the method for generating a representation of a chemical substance, wherein the input data is a multimodal latent space representation of the physicochemical properties of the chemical substance, and the decoded output data is multimodal reconstructed data in a multimodal initial space.

[0150] According to one aspect, a computer program product is provided comprising instructions, which, when executed by a computer, cause the computer to perform the method according to the first aspect or an embodiment of the first aspect.

[0151] The computer program product such as the computer program means may be embodied as a memory card, USB stick, CD-ROM, DVD or a file downloadable from a server in a network. For example, such a file may be provided by transmitting a file comprising the computer program product from a wireless communication network.

[0152] The embodiments and features described with reference to the training method of the first aspect or an embodiment of the first aspect apply mutatis mutandis to the computer program product of the second aspect.

[0153] According to one aspect, a database search device is provided, which is particularly implemented to identify chemical substances having predetermined physicochemical properties. The database search device comprises:

[0154] A storage unit, the storage unit is used to store a database and a trained neural network, the database includes representations of multiple chemical substances, especially in a latent space, these representations are obtained using the trained neural network, and the trained neural network includes an encoder and a decoder;

[0155] an input unit for receiving search input data indicating a chemical substance to be searched;

[0156] A processor configured to:

[0157] adjusting the dimension of the search input data using the encoder to obtain encoded search data in an encoded space;

[0158] comparing the encoded search data to representations of the plurality of chemical substances in the database;

[0159] selecting at least one chemical substance from the plurality of chemical substances represented in the database based on a comparison between the encoded search data and the representation of the plurality of chemical substances; and

[0160] The output unit is used to output the identifier of the selected chemical substance.

[0161] The trained neural network may include a variational autoencoder including a multimodal encoder and a multimodal decoder.

[0162] The database search device may be part of a computer, in particular a personal computer or an industrial computer. The trained neural network may be used to provide a latent space representation of chemical substances. In particular, the database comprises latent space representations of a plurality of chemical substances obtained using the trained neural network. For example, the database may be regularly and / or continuously updated when new data about chemical substances are obtained.

[0163] The storage unit storing the database and the trained neural network may be any type of temporary or persistent storage device (memory). The processor may be a central processing unit (CPU) or the like configured to access the database and / or execute the neural network stored therein. The input unit may include a user interface for receiving search input data from a user, or may be a unit that can access the search input data stored in the storage unit or the like.

[0164] Search input data is data that has not yet been input into the neural network and / or for which no potential representations are stored. The search input data may have the same format as the multimodal input data described previously. The search input data may also be incomplete data representing a chemical substance, for example, including representations of only some of the modalities of the chemical substance.

[0165] The neural network can be used to bring the search input data into the same latent space representation as the data in the database. This is performed using a multimodal encoder of the neural network that is particularly capable of combining multiple modalities of the search input data.

[0166] Preferably, the potential search data has the same representation (and dimensionality) as the data in the database. Comparison of the potential search data with the data in the database (i.e., with the representations of the plurality of chemical substances) can be performed by directly comparing the potential search data with the data in the database. For example, the values ​​of the potential search data assigned to each of its dimensions can be directly compared with the values ​​of each data in the database assigned to the same dimension. This comparison allows the determination of the similarity between the potential search data and each representation of the plurality of chemical substances in the latent space. A comparison score proportional to the degree of similarity can be assigned to each representation of the plurality of chemical substances in the latent space.

[0167] At least one of the selected chemicals can be the chemical whose latent space representation in the database is closest to the potential search data (most similar, e.g., has the highest comparison score). The plurality of selected chemicals can be the N chemicals represented in the database that are closest to the potential search data. This relies on the fact that similar chemicals will have similar latent space representations.

[0168] The output unit may be a user interface, such as a display, a touch screen, etc. The identifier of the selected chemical substance may include the name, chemical composition, reference number or other identification information of the selected chemical substance. The output unit may output the identifier by outputting (displaying) the identifier to a user, storing it in a storage unit, etc.

[0169] The database search device can be used to identify chemicals based on multimodal representations of chemicals (search input data) by performing comparisons in a latent space.

[0170] The database search device can be used to perform redesign, that is, find the representation of the chemical substance without knowing the formula of the chemical substance. In this case, once the unknown chemical substance search result is searched, its formula can be derived from the formula of the selected chemical substance.

[0171] Furthermore, the latent space representation provided by neural networks can alleviate the few-shot learning problem (the problem of making predictions based on a limited number of samples) by reducing the dimensionality of the input data and by providing a rich feature space that has been trained on very large datasets.

[0172] According to an embodiment, the neural network is trained according to the method of the first aspect or any embodiment thereof.

[0173] The database search device may be further configured to perform training of the neural network according to the method of the first aspect or any embodiment thereof.

[0174] Another aspect of the present disclosure relates to a method for generating control data indicative of synthesis specifications of a chemical substance, particularly a polymer, the method comprising:

[0175] Provide first synthesis specifications for reference chemicals;

[0176] encoding the first synthesis specification into a digital representation of the reference chemical substance using a data driven compression model;

[0177] providing a database including a plurality of historical digital representations of historical chemical substances;

[0178] determining a similarity score of the historical digital representations relative to the digital representation of the reference chemical;

[0179] Based on the similarity score, selecting at least one historical representation, and decoding a synthesis specification associated with the at least one selected historical representation; and

[0180] Control data indicating the generated synthesis specification is generated.

[0181] The synthesis specification of a chemical substance preferably includes all process and recipe data required to produce the corresponding chemical substance.The synthesis specification data may include control data in machine-readable form for operating a chemical facility.

[0182] As a result of the above aspects, if the control data is deployed in a chemical facility (i.e., in a system for producing a chemical substance according to the control data), synthesis specifications that can produce a chemical substance similar to the reference chemical substance can be obtained. The method also provides alternative synthesis specifications for the reference substance.

[0183] In an embodiment, the data-driven compression model is implemented to map input data to encoded output data according to a first aspect of a method for generating a representation of a chemical substance, wherein the input data is a multimodal representation of the chemical substance and the encoded output data is a latent space representation of the input data.

[0184] The database may be included in the database search device according to the aforementioned aspect.

[0185] In an embodiment, the search input data indicates a component of the chemical substance to be replaced by an alternative component, and the selected chemical substance includes the alternative component instead of the component to be replaced.

[0186] In an embodiment, the search input data indicates a qualitative property of the chemical substance. The quality may refer to a classification according to a predetermined regulation, for example according to the German Hazardous Substances Ordinance (Gefahrstoffverordnung) - GefStoffV.

[0187] In particular, the identifier indicating the component and / or qualitative property to be replaced may be a modality on the encoder and decoder side of the neural network.

[0188] One or more aspects relate to a method for generating control data indicative of synthesis specifications of a chemical substance, in particular a polymer, the method comprising:

[0189] receiving sensor data indicative of a measurable physicochemical property of the chemical substance;

[0190] encoding the sensor data using a data-driven compression model of the chemical substance to generate encoded sensor data; and

[0191] generating control data indicative of synthesis specifications for the chemical substance by decoding the encoded sensor data using the data-driven compression model;

[0192] Therein, the data-driven compression model is implemented to map input data to encoded output data according to a first aspect of a method for generating a representation of a chemical substance, wherein the input data is a multimodal representation of the physicochemical properties of the chemical substance, and the encoded output data is particularly a latent space representation of the input data.

[0193] The proposed aspects allow the generation of control data for synthesizing chemical substances according to desired physicochemical properties without the need for actual experiments or test runs of chemical facilities. Thus, digital representations in the form of latent space representations facilitate the manufacture and operation of facilities for producing chemicals.

[0194] In an embodiment, the method for generating control data may comprise at least one of the following steps:

[0195] Apply constraints indicating process or substance requirements, in particular requirements for the biodegradability of the chemical substance and / or the ingredients used to produce it, requirements to be biomass-based, requirements to exclude toxic ingredients, etc.; and

[0196] In particular, it is verified that the generated synthesis specifications satisfy the constraints before generating the control data.

[0197] The embodiments and features of the training method aspects described with reference to the first aspect or its embodiments are applicable, with appropriate modifications, to the database device of the third aspect and other method aspects for determining physicochemical properties, generating control data and / or measurement data.

[0198] According to some aspects, the processor is further capable of decoding the latent representation of at least one selected chemical substance using a multimodal decoder to obtain reconstructed representation data in the multimodal initial space; and the output unit is configured to output the reconstructed representation data. Thereby, a representation of the selected chemical substance is provided in the initial space, which is understandable and analyzable by a user, for example.

[0199] The disclosed aspects specifically allow for the substitution of known polymers with molecules having similar performance, including appropriate synthetic specifications for generating alternative polymers / molecules.

[0200] All disclosed methods are preferably computer-implemented.

[0201] Further possible implementations or alternative solutions of the present invention also encompass combinations of features described above or below with respect to these embodiments (not explicitly mentioned herein). A person skilled in the art may also add separate or isolated aspects and features to the most basic form of the present invention.

[0202] Other embodiments, features and advantages of the present invention will become apparent from the following description and the dependent claims, taken in conjunction with the accompanying drawings, in which:

[0203] Figure 1 A first example of a neural network is shown;

[0204] Figure 2 Shown for training Figure 1 A first embodiment of a neural network method;

[0205] Figure 3 Shown for training Figure 1 A second embodiment of the neural network method;

[0206] Figure 4 A second example of a neural network is shown;

[0207] Figure 5 Shown for training Figure 4 A first embodiment of a neural network method;

[0208] Figure 6 Shows Figure 5 Different representations of training methods;

[0209] Figure 7 A database search device is shown;

[0210] Figure 8 Shows the operation Figure 7 A method for searching a database of a device;

[0211] Fig. 9 Shown for use Figure 7 A user interface for a database search device;

[0212] Fig.10 An embodiment of a system for producing a chemical substance is shown; and

[0213] Fig.11 A user interface for using the measurement device / service is shown.

[0214] In the drawings, like reference numbers refer to identical or functionally equivalent elements unless otherwise indicated.

[0215] The encoder, decoder, and autoencoder proposed in this paper can be implemented according to the following paper and the citations therein: M. Wu, N. Goodman: “Multimodal Generative Models for Scalable Weakly-Supervised Learning”, arXiv:1802.05335, which is incorporated herein by reference.

[0216] Figure 1 An example of a neural network 1 is shown, which includes a multimodal variational autoencoder 3 having a multimodal encoder 4 and a multimodal decoder 5. The neural network 1 is based on Figure 2 The method is trained by Figure 1 and Figure 2 .

[0217] It should be understood that, with respect to the following embodiments, the neural network presented implements a framework for digital representation of chemical substances. The deployed artificial neural network can be characterized by its parameters, such as the number and characteristics of neurons implemented, weights, nodes, connections and other configuration parameters. In the context of the present application, the expression "latent space representation" refers to the digital representation of chemical substances (such as polymers below). The "modality" describing chemical substances refers to the physicochemical properties of chemical substances, which are observable by measurement and can be represented in a digital or computer-processable manner, such as spectral representations, rheological representations, thermal representations, chemical representations, structural representations, solubility representations, dispersion representations, viscosity representations and / or surface tension representations of the chemical substances.

[0218] The latent space representation of a chemical substance is compressed relative to the amount of data required for multimodal data including these modalities. For example, the characterization of a chemical substance with an original parameter set describing multiple physicochemical properties and names (e.g., CAS (Chemical Abstracts Service) numbers) can be considered as a multimodal representation requiring several data structures. After generating the latent space representation, latent space data structures indicating the characterization of the same substance can be obtained, wherein the latent space representation requires fewer and / or smaller data structures. For example, the dimensionality of the latent space representation is less than the dimensionality of the initial multimodal representation. Because the encoder and decoder are trained with multimodal material data, the potential information loss caused by encoding is reduced or negligible.

[0219] The multimodal encoder 4 and the multimodal decoder 5 form an interface to the latent space representation 6 and are therefore computer-implemented embodiments of a data-driven compression model.

[0220] To train the neural network 1, the neural network 1 receives multimodal input data 2 as input ( Figure 2 Step S1). Figure 1 In the example of , the multimodal input data 2 includes data of seven modes 2a to 2g representing the same chemical substance (here a polymer). Data 2 can be understood as a predetermined multimodal representation of a polymer. Reference numeral 2 indicates a predetermined multimodal representation including a plurality of seven modes 2a to 2g. Mode 2a includes spectral data from spectral measurements, mode 2b includes rheological data, mode 2c includes X-ray diffraction data, mode 2d includes solubility data, mode 2e includes dispersed clay data (which indicates the interaction of the chemical substance with the clay layered structure), mode 2f includes surface tension data, and mode 2g includes viscosity data for the polymer. Data from all dimensions are obtained by performing corresponding measurements on the polymer using sensors. The neural network 1 receives multimodal input data 2 involving multiple polymers.

[0221] The multimodal input data 2 is provided in an initial space. In the initial space, the data from each modality 2a to 2g has its own dimension, which corresponds here to the dimension of the data sensed by the sensor. Thus, the modalities 2a to 2c have a higher dimension (between 5 and 100) than the modalities 2d to 2g, which are scalars (having only one dimension). Alternatively, in the initial space, the data from each modality 2a to 2g has the same dimension (e.g., 50).

[0222] In an alternative embodiment, encoder 4 is replaced by a plurality of separate encoders, each associated with one of the input modalities 2a to 2g, and decoder 5 is replaced by a plurality of separate decoders, each associated with one of the output modalities 7a to 7g.

[0223] exist Figure 2 In step S2, a multimodal encoder 4 is used to modify the dimensionality of the multimodal input data 2. The multimodal encoder 4 comprises a plurality of multimodal encoder layers, each multimodal encoder layer having multimodal encoder weights defining a mathematical operation according to which the multimodal encoder 4 transforms the multimodal input data 2. The multimodal encoder weights are some of the parameters that are modified and optimized during the training of the neural network 1, as will be further explained below.

[0224] In step S2, the multimodal encoder 4 reduces the number of dimensions of the multimodal input data 2 to obtain multimodal latent data in a latent space 6. In this example, the latent space representation of the multimodal input data 2 (ie, the multimodal latent data) includes 16 dimensions.

[0225] In step S2 , the multimodal encoder 4 combines the data from all modalities 2 a to 2 g to form a single dataset describing the polymer in a latent space 6 .

[0226] In step S3, the multimodal latent data is decoded using a multimodal decoder 5 to obtain multimodal reconstructed data 7 in the initial space, namely 7a to 7g. This includes modifying the dimensionality of the multimodal latent data to return to the dimensionality of the initial multimodal input data 2. The multimodal decoder 5 includes a plurality of multimodal decoder layers, each having multimodal decoder weights defining a mathematical operation according to which the multimodal decoder 5 decodes the multimodal latent data. The multimodal decoder weights are some of the parameters that are modified and optimized during the training of the neural network 1, as will be further explained below.

[0227] exist Figure 2 In step S4 of the training method, a loss function of the multimodal variational autoencoder 3 is calculated. The loss function in its simplest form indicates the similarity between the multimodal input data 2 and the multimodal reconstruction data 7. Alternative ways of calculating the loss function of the multimodal variational autoencoder 3 include mixture of experts, Gaussian mixture and / or expert multiplication techniques.

[0228] The calculated loss function indicates how well the neural network 1 performs during the current run (iteration). The smaller the loss function, the better the neural network 1 is.

[0229] Figure 3 Shown for training Figure 1 A further embodiment of the method of the neural network 1. Figure 3 The method steps S1 to S4 are Figure 2 The steps are the same as in the previous example. Depending on the loss function being calculated, the neural network can Figure 3 In the optional step S5 of , all or some of the multimodal encoder weights are updated and / or all or some of the multimodal decoder weights are updated. These multimodal encoder weights and / or multimodal decoder weights are updated by back propagation.

[0230] like Figure 3 As shown, in step S6, all method steps S1 to S5 can be repeated to reduce the loss function and thus improve the neural network 1. Steps S1 to S5 can be repeated for a predetermined number of runs, or until the calculated loss function is less than a predetermined loss function threshold. When training stops, the multimodal encoder weights and multimodal decoder weights of the run that provides the lowest loss function are kept as the weights that produce the best neural network 1. The trained neural network 1 corresponds to the sub-optimal run and has its multimodal encoder weights and decoder weights.

[0231] Figure 4A second example of a neural network 1 is shown. Figure 5 Shown for training Figure 4 Embodiment of the method of neural network 1. Figure 4 Neural network 1 and Figure 5 Many elements of the method are related to Figures 1 to 3 The neural network 1 is trained in the same way and is also applicable to Figure 4 and Figure 5 Description.

[0232] and Figure 1 The difference of neural network 1 is that Figure 4 The neural network includes seven separate variational autoencoders 10, each of which includes a separate encoder 8 and a separate decoder 9. In detail, the separate encoders 8a to 8g and the separate decoders 9a to 9g correspond to the modalities 2a to 2g, respectively. The modalities 2a to 2g correspond to the modalities 2a to 2g described previously, but their representation data form separate input data 12 instead of the multimodal input data 2. The difference between the separate input data 12 and the multimodal input data 2 is that the separate input data 12 is input to the separate variational autoencoder 10, while the multimodal input data 2 is input to the multimodal variational autoencoder 3. Further, the separate input data 12 includes data of different dimensions for different modalities, while the multimodal input data 2 may include data of the same dimension for all modalities 2a to 2g.

[0233] The individual variational autoencoders 10 are used to condition the data 12 before inputting it into the multimodal variational autoencoder 3. The individual encoders 8a to 8g bring the input data 12 from each modality 2a to 2g into the same predetermined dimension, which may be the dimension of the latent space 6 (e.g., dimension 16).

[0234] In an embodiment, the autoencoder 3 is an optional element, and separate encoders 8a to 8g each bring the input data 12 from each modality 2a-2g into a predetermined dimension of the same latent space 6. Likewise, separate decoders 9a to 9g map the latent space vectors to the corresponding modality 17a to 17g having a specific individual dimension.

[0235] The previous training makes various dimensions and modalities closely related, so that the individual decoders / encoders 8, 9 interact with latent space vectors with predetermined dimensions. The missing input modality can be repaired by the autoencoder structure.

[0236] In detail, Figure 5In step S6, the individual input data 12 representing each individual modality 2a to 2g is input to the corresponding individual encoder 8. This means that the individual input data 12 representing modality 2a is input to the corresponding individual encoder 8a, the individual input data 12 representing modality 2b is input to the corresponding individual encoder 8b, and so on.

[0237] exist Figure 5 In step S7, each individual encoder 8 modifies the dimension of the received individual input data 12 to obtain data with a predetermined dimension (e.g., 16). The data with a predetermined dimension obtained is called "individual latent data" and may correspond to Figure 1 Description of the multimodal input data2.

[0238] In step S8, the separate latent data are reconstructed using the separate decoders 9a to 9g to obtain separate reconstructed data 17 in the separate initial space (i.e., in the same space as the separate input data 12). The separate reconstructed data 17 includes separate data 17a to 17g for each modality 2a to 2g. The separate reconstructed data 17 may be Figure 1 The multimodal reconstructed data 7 in are located in the same space and are identical to it, or they can be located in different spaces (separate latent spaces).

[0239] In step S9, the individual input data 12 from each modality 2a to 2g is compared with the corresponding individual reconstruction data 17a to 17g to obtain a comparison result. The better each individual variational autoencoder 10 is, the more similar its input data 12 is to the reconstruction data 17. The comparison result can be a loss function.

[0240] Accordingly, in Figure 5 In step S10, the weight of the variational self-encoder 10 is updated according to the comparison result of the individual variational self-encoder. In particular, based on the comparison result obtained by comparing the input data 12 of the modality 2a with the individual reconstruction data 17a, the individual encoder weights of the individual encoder 8a and the individual decoder weights of the individual decoder are updated by back propagation. The same operation is performed for each individual variational self-encoder 10.

[0241] like Figure 5 As shown, in step S21, the steps of training the individual variational self-encoder 10 (steps S6 to S10) are repeated to reduce the comparison result and thus improve the individual variational self-encoder 10. Steps S6 to S10 may be repeated until the desired comparison result is achieved, or until a predetermined number of times or runs are performed.

[0242] exist Figure 5In step S11, the individual latent data of the trained variational autoencoder 10 is used as Figures 1 to 3 The multimodal input data 2 of the multimodal variational self-encoder 3 described. After step S11, Figure 5 The method performs method steps S1 to S4, wherein separate latent data of the trained variational autoencoder 10 is used as multimodal input data 2 of the multimodal variational autoencoder 3.

[0243] Figure 6 Another representation of the training procedure of the neural network 1 is shown. Figure 6 Boxes 13, 14 and 15 in the figure represent model selection 13, individual optimization 14 and hyperparameter optimization 15 respectively.

[0244] In step S22, individual input data 12 of the modalities 2a to 2g are collected. Steps S23 to S25 are part of an individual optimization and include considering Figure 5 Steps S6 to S11 described above. In step S24, a search space for hyperparameters of a separate variational autoencoder 10 is defined (this includes weights, number of layers, activation functions, channel size, etc.). In step S25, the architecture of the separate variational autoencoder 10 is optimized, in particular consistent with steps S6 to S11. Step S23 indicates that steps S24 and S25 are performed for each modality 2a to 2g. The result of steps S23 to S25 (i.e., the output of the separate optimization 14) is a variational autoencoder 10 optimized for each modality 2a to 2g.

[0245] This output is used as input to step S26, in which the multimodal variational autoencoder 3 is trained for the fixed model architecture defined in steps S23 to S25. Step S26 may include steps S1 to S4 previously defined. Step S26 may include optimization of the hyperparameters of the latent space. Thus, a joint representation of all modalities 2a to 2g in the latent space 6 is obtained. The optimization in steps S25 and S26 is a Bayesian optimization of the hyperparameters.

[0246] Arrow 16 indicates that steps S23 to S26 are repeated for different values ​​of the predetermined dimension in order to optimize the loss function of the multimodal variational autoencoder 3 and achieve an optimal latent space representation of the chemical substance.

[0247] The hyperparameters that minimize the loss function are saved in step S27. In particular, all information related to the trained and optimized neural network is stored. This includes the latent space variables for each data set, together with information about modalities 2a to 2g and all further available information. Figure 6 In step S28, the trained neural network 1 is used to run the application test.

[0248] By considering Figures 1 to 5 The training method described provides a neural network 1 that is capable of representing a polymer in a latent space representation. In detail, training data and further data representing the polymer can be input into the trained neural network. The trained neural network generates a latent representation of the input data, which can be stored in a database. This allows for a variety of applications, which will be described in detail below.

[0249] An example of an application of the trained neural network 1 is a database search device 20 (search engine). An example of such a database search device 20 is Figure 7 Shown in.

[0250] The search device may implement various functions and support various methods, for example, a method for generating control data indicating synthesis specifications of a desired chemical substance, or a method for generating synthesis measurement data.

[0251] Figure 7 The database search device 20 includes a storage unit 21 as a random access memory (RAM), an input unit 23, a processor 24 as a CPU, an output unit 25, and a connection cable 26 that connects the different components of the database search device 20.

[0252] The database search device 20 is a part of a personal computer (PC). The storage unit 21 has a database 22 and a trained neural network 1 stored thereon. The database 22 includes latent space representations of a plurality of chemical substances (e.g., polymers) obtained from the trained neural network 1. In detail, in order to obtain the latent space representation stored in the database 22, the trained neural network 1 receives the separate input data 12 and / or the multimodal input data 2 previously used as training data, and generates a latent space representation in the latent space 6 using the multimodal variational autoencoder 3 and / or the separate variational autoencoder 10.

[0253] Figure 8 An example of how to use the database search device 20 is shown and will be described below in conjunction with Figure 7 and Figure 8 The database search device 20 is used to search the database 22 to obtain polymers that are the same as or similar to the polymer being searched. Fig. 9 A user interface 31 of the database search device 20 is shown.

[0254] exist Figure 8In step S12 of the method, the input unit 23 receives search input data providing a multimodal representation of the polymer to be searched. The search input data is provided in a multimodal initial space. The search input data has the same format as the previously described multimodal input data 2, wherein the data describes a plurality of modes 2a to 2g of the polymer. Optionally, the search input data only includes data describing some of the modes 2a to 2g.

[0255] The input portion 32 of the user interface 31 has a drop-down menu 34 and an input field 35 into which the user can insert the multimodal data 2 . Fig. 9 The following potential modalities are shown: CAS number, density, pH, specific NMR data (can be uploaded), and viscosity. For example, alternatives to C12-15 branched and straight chain alcohols are desired. Fig. 9 In the Example view, the ethoxylated propoxylate corresponding to CAS 1755111905-53-4 is entered together with accessible physicochemical properties (density, pH viscosity and NMR file).

[0256] As described above, a latent space representation of the multimodal substance data 2 input via the interface 32 is generated by the processor 24 according to the above method. Within the latent space representation, similar chemical substances are searched, for example by finding latent space vectors in a similarity region around the latent space vector corresponding to the input substance ethoxylated propoxylate.

[0257] Fig. 9 The right side of shows the search results: As an alternative to the ethoxylated propoxylates, butoxylated ethoxylates are proposed, namely C13-15-branched and linear alcohols corresponding to CAS 120313-48-6, which have the indicated physicochemical properties.

[0258] The interface may also output other modalities of desired input material, such as recipes or control data for a chemical reactor.

[0259] In another example, the search input data includes the physicochemical properties of the desired chemical substance, such as specific thermal conductivity. Therefore, the method implemented by the database search device 20 outputs control data indicating synthesis specifications. The control data is suitable for specifying the required elements of the chemical facility and controlling those elements to produce the chemical substance (polymer in the described example). The control data may include a digital version of a recipe for producing a chemical substance with the desired properties.

[0260] In step S13, the search input data is transformed into its latent space representation using the processor 24. In detail, the multimodal encoder 4 of the neural network 1 is used to adjust the dimensionality of the search input data to obtain multimodal latent search data in the latent space 6. Therefore, by deploying the data-driven compression model implemented by the encoder 4 and the decoder 5, a digital representation of the chemical substance (e.g., polymer) is obtained.

[0261] In step S14, the processor retrieves the latent space representation of previously known polymers from the database 22 stored in the storage unit 21. The database 22 may include latent space representations of historical or known polymers.

[0262] exist Figure 8 In step S15, processor 24 compares the potential search data from step S13 with representations of a plurality of polymers retrieved from database 22. This may involve calculating a similarity score.

[0263] exist Figure 8 In step S16, processor 24 selects at least one polymer from the plurality of polymers represented in database 22 based on the comparison result of step S15. The scores may be sorted in step S17 so that a list of similar or close polymers in the latent space may be obtained for further selection.

[0264] Here, processor 24 selects the closest polymer(s) in latent space 6 (eg, the closest Euclidean distance between points representing polymers in latent space 6 ).

[0265] In an optional step S17, the selected closest polymers are sorted by distance (ie according to their similarity to the potential search data).

[0266] exist Figure 8 In step S18, an identifier is obtained from the database 22, the identifier including information on analytical data, polymer name, synthesis specifications, etc. related to the selected polymer(s).

[0267] In step S19, the output unit 25 as a display outputs the identifier of the selected polymer(s). The identifier is also stored in the database 22. In step S20, the output identifier and / or its associated synthesis specification is used to control the synthesis of the new (searched) polymer. The identifier allows to select a polymer from the specification database 530 (see Fig.11 ) to obtain the specified synthesis specifications associated with the identified polymer. Step S20 may involve running an application test.

[0268] Fig.10A system 500 for producing chemical substances based on synthesis specifications generated according to the above aspects and embodiments of the method and apparatus for generating control data is shown. In this example, the system includes a user interface 510 and a processor 520 associated with a control unit 540, which is configured to receive control data generated according to the present disclosure. In this example, the control data is provided from a database 530, and in other examples, the control data can be provided from a server. For example, an identifier of a specific set of control data is obtained according to step S18, wherein the identifier refers to its associated synthesis specification and corresponding control data set.

[0269] Container 550,552 comprises the component of chemical product separately.Generally can there be more than two containers.For the purpose of explanation, this example only shows two containers.Valves 560,562 are associated with containers 550,552.According to synthesis specification, valves 550 and 552 can be controlled to use appropriate amount of each component as the composition for synthesizing selected polymer (step S17) in reactor 570.The motor 600 of mixer 580 can also be controlled by control unit 540 according to control data / synthesis specification.Optional heater 590 can also be controlled according to synthesis specification.Finally, the control unit can control the outlet valve 610 that is communicated with reactor fluid to provide chemical product to container or test system 620.

[0270] Fig.11 Another example of a user interface 41 that can be used to access a computer-implemented method for measuring physicochemical properties of a chemical substance is shown in FIG. 4. In this example, a density measurement is desired for CAS 1755111905-53-4C13-15 alcohol, but only information about pH, viscosity, and NMR data is available and entered in sections 44, 45. A processing device (e.g., Figure 7 The processor 24 receives the input multimodal data (see Figure 1 The trained neural network is deployed as described above to generate the encoded latent space data.

[0271] The processor generates measurement data indicative of the desired measurable physicochemical properties (density) of the chemical substance (C13-15-branched and linear butoxylated ethoxylated alcohols) by decoding the encoded sensor data using a data driven compression model. This is output in an output portion 43. Thus, the data driven model reconstructs the modes (density) that are missing as input based on the input. Thus, the measurement data can be obtained indirectly through a trained model / neural network (particularly using a variational autoencoder device as described in the present disclosure).

[0272] Although the present invention has been described according to preferred embodiments, it is obvious to those skilled in the art that all embodiments may be modified. For example, the modalities 2a to 2g may be other modalities in addition to the above-mentioned modalities. The neural network 1 may be used for other applications in addition to the above-mentioned database search device 20. Such applications include, for example, polymer redesign based on output identifiers, polymer synthesis based on output identifiers, new polymer design based on output identifiers, reduced sample learning, etc.

[0273] In alternative embodiments and applications of the trained autoencoder or database search device, synthetic measurement data of the chemical substance is obtained based on the available sensor data and the underlying data-driven compression model. It is also conceivable to generate data indicative of a second physicochemical property based on a second physicochemical property, wherein the first property and the second property are related to different modalities with respect to the multimodal latent space representation.

[0274] It should be understood that all disclosed methods can be implemented as computer-implemented methods. In all methods involving generating control data, the optional step of producing chemical substances according to synthesis specifications and / or control data can be performed by using a system for producing chemical substances having a control unit.

[0275] Reference numerals

[0276] 1 Neural Network

[0277] 2 Multimodal Input Data

[0278] 2a to 2g modes

[0279] 3 Multimodal Variational Autoencoder

[0280] 4 Multimodal Encoder

[0281] 5 Multimodal Decoder

[0282] 6 Latent Space

[0283] 7 Multimodal reconstruction data

[0284] 7a to 7g Multimodal reconstruction data

[0285] 8 Separate encoders

[0286] 8a to 8g Separate encoder

[0287] 9 Separate decoder

[0288] 9a to 9g Separate decoder

[0289] 10 Single Variational Autoencoder

[0290] 12 Separate input data

[0291] 13 Model selection

[0292] 14 Individual Optimization

[0293] 15 Hyperparameter Optimization

[0294] 16 Arrow

[0295] 17 Separate reconstruction data

[0296] 17a to 17g separate reconstruction data

[0297] 20 Database search device

[0298] 21 Storage Unit

[0299] 22 Database

[0300] 23 Input unit

[0301] 24 Processors

[0302] 25 Output Units

[0303] 26 Connecting cables

[0304] 31 User Interface

[0305] 32 Input section

[0306] 33 Output section

[0307] 34 Drop-down menu

[0308] 35 Input Field

[0309] 41 User Interface

[0310] 42 Input Section

[0311] 43 Output section

[0312] 44 Drop-down menu

[0313] 45 Input Field

[0314] 500 Systems / chemical facilities for the production of chemical substances

[0315] 510 interface

[0316] 520 processor

[0317] 530 Database

[0318] 540 Control Unit

[0319] 550, 552 containers

[0320] 560, 562 valve

[0321] 570 Components / ingredients

[0322] 580 Mixer

[0323] 590 Heater

[0324] 600 Motor

[0325] 610 Outlet valve

[0326] S1 receives multimodal input data

[0327] S2 adjusts / reduces multimodality by encoding based on a data-driven compression model

[0328] Dimensionality of input data

[0329] S3 decoding based on data-driven compression model

[0330] S4 Calculate the loss function

[0331] S5 Update codec weights

[0332] S6 Repeat steps S1 to S5

[0333] S7 Adjust Dimensions

[0334] S8 decodes individual latent data based on a data-driven compression model

[0335] S9 compares the individual latent data with the reconstructed data

[0336] S10 Update codec weights

[0337] S11 Using separate latent data as multimodal input data

[0338] S12 Receive multimodal representation / measurement data

[0339] S13 Encoding search input data / transformation into latent space

[0340] S14 Obtaining latent space representations of known / historical chemicals

[0341] S15 Compare the search data with the known data in the latent space / determine the latent space

[0342] The distance between

[0343] S16 determines the closest match between the search chemical and known chemicals in the latent space

[0344] Close point / select material

[0345] S17 searches for the potential space table closest to the input data based on the potential space distance.

[0346] Sort / select a group of chemicals shown

[0347] S18 Get the identifier of the selected / closest chemical substance

[0348] S19 Acquire / generate control data indicating the synthesis specifications of the closest chemical substance

[0349] S20 Controls the synthesis / run test application of selected chemicals according to synthesis specifications

[0350] S21 Repeat steps S6 to S10

[0351] S22 receives multi-modal input data

[0352] S23 Execute steps for each mode

[0353] S24 sets the search space for a single autoencoder

[0354] S25 Optimizing the architecture of variational autoencoders

[0355] S26 Training Variational Autoencoder

[0356] S27 stores hyperparameters

[0357] S28 Run application test

Claims

1. A method for characterizing a chemical substance in a predetermined multimodal representation, the predetermined multimodal representation having a predetermined plurality of modalities, the method comprising: Receiving (S12) multimodal material data (2, 12) including a first set of modalities of the chemical substance; encoding the multimodal data (2, 12) using a data-driven model (3, 10) of the chemical substance (S13) to generate encoded substance data; as well as Generating multimodal substance data comprising a second set of modalities of the chemical substance by decoding the encoded substance data using the data-driven model (3, 10), wherein the multimodal substance data indicates physicochemical properties of the chemical substance, a composition of the chemical substance and / or an identifier of the chemical substance; wherein the data-driven model (3, 10) is implemented to map input data (2, 12) to encoded output data (6), the input data being a multimodal representation of the chemical substance (2, 12), and the encoded output data being a latent space representation (6) of the input data (2, 12); wherein the first group of modalities is different from the second group of modalities, and wherein the first group of modalities and the second group of modalities are included in the predetermined plurality of modalities, wherein, in particular, at least one modality of the second group is not included in the first group.

2. The method of claim 1, wherein: The first set of multiple modalities is a subset of the second set of multiple modalities.

3. The method according to claim 1 or 2, wherein: The data-driven model comprises at least one trained neural network (1), which is implemented to receive multimodal input data (2, 12) comprising the predetermined multiple modalities, encode the input data into a latent space representation (6) of the input data (2, 12), and decode the encoded input data into multimodal output data comprising the predetermined multiple modalities.

4. The method of claim 3, wherein: The neural network (1) is trained based on training data, wherein the training data includes multimodal training data, and the multimodal training data includes the predetermined multiple modalities.

5. The method according to claim 3 or 4, wherein: The neural network (1) comprises a plurality of individual encoders (8a to 8g), wherein each individual encoder (8a to 8g) is assigned to one modality (2a to 2g) of the predetermined plurality of modalities, wherein each individual encoder (8a to 8g) is implemented to bring input data (2, 12) from the modality (2a to 2g) to which the individual encoder (8a to 8g) is assigned into the same dimension of the latent space (6).

6. The method according to any one of claims 3 to 5, wherein: The neural network (1) comprises a plurality of individual decoders (9a to 9g), wherein each individual decoder (9a to 9g) is assigned to one modality (7a to 7g) of the predetermined plurality of modalities, wherein each individual decoder (9a to 9g) is implemented to decode the latent space representation (6) of the encoded input data (2, 12) into modal data in the generated multimodal material data, which modality is the modal data of the modality to which the individual decoder (8a to 8g) is assigned.

7. The method according to any one of claims 1 to 6, wherein: Characterization includes measuring the physicochemical properties of the chemical substance. The substance data includes sensor data. The measurements include: receiving (S12) sensor data (2, 12) indicative of a first measurable physicochemical property of the chemical substance, the sensor data being associated with at least one modality included in the first group; encoding the sensor data (2, 12) using a data-driven compression model (3, 10) of the chemical substance (S13) to generate encoded sensor data; and Measurement data indicative of a second measurable physicochemical property of the chemical substance is generated by decoding the encoded sensor data using the data driven compression model (3, 10), the measurement data being associated with at least one modality included in the second group.

8. The method of claim 7, wherein: At least one of the group consisting of: the generated measurement data, the formulation data indicating the chemical substance, and the identification data indicating the chemical substance is output.

9. The method according to any one of claims 1 to 8, further comprising: For a plurality of sample chemicals, using the data-driven compression model to generate a latent space representation (6) of multimodal sample material data associated with the sample chemicals, the multimodal sample material data having the predetermined plurality of modalities; and / or The generated latent space representation (6) of the multimodal sample material data associated with the sample chemical substances is stored in a database.

10. The method according to any one of claims 1 to 9, further comprising: receiving search input data indicating predetermined physicochemical properties of a chemical substance to be searched, wherein the search input data is provided as a multimodal representation of the substance to be searched included in the first group; encoding received search input data to generate a latent space representation of the search input data; and The generated latent space representation of the search input data is compared with the latent space representation of the sample chemical substances to obtain a comparison result.

11. The method according to claim 9 or 10, further comprising: In response to the comparison, at least one sample chemical substance is selected.

12. The method according to any one of claims 9 to 11, wherein: Comparisons include: calculating a similarity score of the latent space representation of the search input data relative to the latent space representation of the sample chemical; and / or A similarity range within the latent space relative to the latent space representation of the search input data is determined.

13. The method according to any one of claims 1 to 12, wherein: At least one modality in the first group and / or the second group includes synthesis specifications of the chemical substance, and / or control data indicative of synthesis specifications of the chemical substance.

14. The method according to any one of claims 1 to 13, wherein: Characterization includes the generation of control data indicative of the synthesis specifications of the chemical substance, particularly the polymer, including: providing a first synthesis specification of a reference chemical substance as at least one modality in the first group; encoding (S13) the first synthesis specification as a digital representation (6) of the reference chemical substance using the data driven compression model (3, 10); providing a database (22) comprising a plurality of historical digital representations of historical chemical substances; determining a similarity score of the historical digital representations relative to the digital representation of the reference chemical substance (S16); Based on the similarity score, selecting (S17) at least one historical representation, and decoding and generating a synthesis specification associated with the at least one selected historical representation; and Control data indicating the generated synthesis specification is generated.

15. The method according to claims 9 and 14, wherein: These historical chemicals are sample chemicals.

16. The method according to any one of claims 1 to 15, wherein: Characterization involves the generation of control data that indicates the synthesis specifications of chemical substances, particularly polymers, including: receiving (S12) sensor data (2, 12) indicative of measurable physicochemical properties of the chemical substance; encoding the sensor data (2, 12) using a data-driven compression model (3, 10) of the chemical substance (S13) to generate encoded sensor data; and Control data indicating synthesis specifications of the chemical substance is generated by decoding the encoded sensor data using the data driven compression model (3, 10).

17. The method according to any one of claims 9 to 16, wherein: The first set of modes is equal to the second set of modes.

18. The method according to any one of claims 1 to 17, wherein: Using the data driven compression model includes a process for generating a compressed digital representation of a chemical substance, particularly a polymer, including: receiving (S12) input data that is a multimodal representation (2, 12) of a physicochemical property of the chemical substance and is indicative of a measurable physicochemical property of the chemical substance; encoding the input data (2, 12) using a data-driven compression model (3, 10) of the chemical substance (S13) so as to generate encoded substance data according to the received input data (2, 12); and Chemical substance data indicating measurable physicochemical properties of the chemical substance is generated by decoding the encoded substance data using the data driven compression model (3, 10).

19. The method of claim 18, wherein: The process for generating the compressed digital representation further comprises: Providing (S1) multimodal input data (2) as input to a neural network (1) to be trained, the multimodal input data providing a multimodal representation of the chemical substance in a multimodal initial space, the neural network (1) comprising a multimodal variational autoencoder (3), the multimodal variational autoencoder comprising a multimodal encoder (4) and a multimodal decoder (5), wherein the multimodal encoder (4) comprises multimodal encoder layers having multimodal encoder weights defining how the multimodal encoder layers transform data, and wherein the multimodal decoder (5) comprises multimodal decoder layers having multimodal decoder weights defining how the multimodal decoder layers transform data; Using these multimodal encoder layers, adjusting (S2) the dimensionality of the multimodal input data (2) to obtain multimodal latent data in a latent space (6); Decoding the multimodal latent data (S3) using the multimodal decoder layers to obtain multimodal reconstructed data in the multimodal initial space; Calculating (S4) a loss function of the multimodal variational autoencoder (3) under a current set of multimodal encoder and decoder weights based on the multimodal input data (2) and the multimodal reconstruction data; and outputting configuration data indicative of the trained neural network (1) and / or a latent space representation of the chemical substance as a digital representation of the chemical substance; and The trained neural network is provided as the data-driven compression model.

20. The method of claim 19, wherein: The process for generating the compressed digital representation further comprises: updating (S5) the multimodal encoder weights and / or the multimodal decoder weights based on the loss function; and / or Repeat (S6) the steps of providing (S1) multimodal input data (2), adjusting (S2) the dimensionality, decoding (S3) the multimodal latent data, calculating (S4) the loss function and / or updating (S5) the weights to reduce the loss function.

21. The method of any one of claims 19 or 20, wherein: The neural network (1) further comprises a separate variational autoencoder (10) assigned to each modality (2a to 2g), the separate variational autoencoder (10) comprising a separate encoder (8) and a separate decoder (9), wherein the separate encoder (8) comprises separate encoder layers having separate encoder weights defining how the separate encoder layers transform data, and wherein the separate decoder (9) comprises separate decoder layers having separate decoder weights defining how the separate decoder layers transform data, the method further comprising: Each individual variational autoencoder (10) is trained as follows: Inputting (S6) a single input data (12) representing only the assigned modality in the single initial space into the single variational autoencoder (10); adjusting (S7) the dimensionality of the single input data (12) using the single encoder layer to obtain single latent data having a predetermined dimensionality; Decoding the single latent data (S8) using the single decoder layer to obtain the single reconstructed data (17) in the single initial space; comparing (S9) the single input data (12) with the single reconstructed data (17) to obtain a comparison result; updating (S10) the individual encoder weights and / or the individual decoder weights based on the comparison result; and Repeating (S21) the steps of inputting (S6) separate input data (12), adjusting (S7) the dimension of the separate input data (12), decoding (S8) the separate latent data, comparing (S9) the data, and updating (S10) the separate encoder weight and / or the separate decoder weight to reduce the comparison result; And the method further comprises Using (S11) individual latent data from a plurality of individual variational autoencoders (10) as multimodal input data (2) of the multimodal variational autoencoder (3).

22. The method according to claim 21, further comprising: Repeating the training of each individual variational autoencoder (10) for different predetermined dimensions of the individual latent data; as well as The predetermined dimension with the smallest loss function is selected as the predetermined dimension of the final trained neural network (1).

23. The method of any one of claims 19 to 22, wherein: The loss function of the multimodal variational autoencoder (3) is determined by a mixture of experts or by an expert multiplication technique.

24. The method of any one of claims 19 to 23, wherein: One of the modalities (2a to 2g) describing the chemical substance provides: a spectral representation, a rheological representation, a thermal representation, a chemical representation, a structural representation, a representation of solubility, a representation of dispersion, a representation of viscosity and / or a representation of surface tension of the chemical substance.

25. A measuring device for measuring the physicochemical properties of a chemical substance according to claim 7, the measuring device comprising: an interface device implemented to receive sensor data (2, 12) indicative of a first measurable physicochemical property of the chemical substance; an encoder device (4) implemented to encode received sensor data (2, 12) and to generate and output encoded sensor data (6); as well as a decoder device (5) implemented to generate measurement data (7, 17) indicative of a second measurable physicochemical property of the chemical substance and to decode the encoded sensor data (6); wherein the encoder device (4) is implemented to map input data (2, 12) to encoded output data (6) according to the method of any one of claims 1 to 24, wherein the input data (2, 12) is a multimodal representation of the physicochemical properties of the chemical substance in the multimodal initial space, and the encoded output data (6) is a latent space representation of the input data (2, 12); and Wherein, the decoder device (5) is implemented to map input data (6) to decoded output data (7, 17), in particular according to the method described in any one of claims 1 to 7, wherein the input data (6) is a multimodal latent space representation of the physicochemical properties of the chemical substance, and the decoded output data (7, 17) is multimodal reconstructed data in the multimodal initial space.

26. A database search device (20) for identifying chemical substances having predetermined physicochemical properties according to claim 10, the database search device comprising: A storage unit (21) for storing a database (22) and a trained neural network (1), wherein the database (22) includes representations of a plurality of chemical substances having physicochemical properties, the representations being obtained using the trained neural network (1), and the trained neural network (1) includes an encoder (3) and a decoder (5); an input unit (23) for receiving search input data providing a representation indicative of a predetermined physicochemical property of a chemical substance to be searched for; A processor (24), the processor being configured to: Using the encoder (4) to adjust the dimension of the search input data to obtain encoded search data; comparing the encoded search data to representations of the plurality of chemical substances in the database (22); as well as selecting at least one chemical substance from the plurality of chemical substances represented in the database (22) based on a comparison between the encoded search data and the representation of the plurality of chemical substances; as well as An output unit (25) is used to output an identifier indicating the selected chemical substance and / or a synthesis specification associated with the identifier of the selected chemical substance.