A Machine Learning-Based Method for Predicting the Dissociation Constant of Photoacidogens and an Electronic Device

By using a machine learning-based dual-branch deep learning model, combined with graph neural networks and multilayer perceptrons, the problems of high cost and insufficient prediction accuracy in obtaining the pKa value of photoacid generators are solved. This achieves efficient and accurate prediction of the dissociation constant of photoacid generators, which is suitable for photoresist material design.

CN123120104APending Publication Date: 2026-11-17张江国家实验室
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611407800.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for obtaining the pKa value of photoacid-producing agents are costly and time-consuming, and deep learning methods do not capture complete features when predicting complex fluorine-containing molecules, resulting in insufficient prediction accuracy.

Method used

A machine learning-based dual-branch deep learning model, combining graph neural networks and multilayer perceptrons, is adopted. Molecular graph features and global molecular structure fingerprints are independently encoded, and the model is trained on a small sample dataset using a transfer learning strategy to output the predicted value of the dissociation constant of photoacidogens.

Benefits of technology

It enables efficient and accurate prediction of the dissociation constant of photoacid-producing agents, meets the requirements of high-throughput screening, significantly reduces R&D costs and accelerates the iteration cycle, and improves prediction accuracy and generalization ability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120104A_ABST
    Figure CN123120104A_ABST
Patent Text Reader

Abstract

This application provides a machine learning-based method and electronic device for predicting the dissociation constant of photoacid-producing agents. The method includes: obtaining the SMILES formula of the target photoacid-producing agent or its acid-producing molecule; converting the SMILES formula into a two-dimensional descriptor of the molecule using molecular informatics tools; inputting molecular graph features and molecular global structure fingerprints as input data into a trained dual-branch deep learning model, enabling the model to output a predicted value of the dissociation constant of the target photoacid-producing agent; the dual-branch deep learning model includes a graph neural network branch, a multilayer perceptron branch, a feature fusion layer, and a fully connected layer. The graph neural network branch and the multilayer perceptron branch independently encode the molecular graph features and molecular global structure fingerprint, respectively. The feature fusion layer concatenates the encoding results of the graph neural network branch and the multilayer perceptron branch along the feature dimension to obtain a fused feature vector, and the fully connected layer performs dimensionality reduction processing on the fused feature vector.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application mainly relates to the field of photoresist material design, and in particular to a machine learning-based method for predicting the dissociation constant of photoacid generators and an electronic device. Background Technology

[0002] Photoacid generators (PAGs) are a core component in chemically amplified (CAR) photoresists, directly affecting the precision of lithographic patterns in extreme ultraviolet (EUV) and deep ultraviolet (DUV) lithography. The strength of the acid generated by PAG after exposure, i.e., the dissociation constant (pKa value), directly determines the catalytic deprotection efficiency of the acid, the acid diffusion length (Blur), and the final lithographic resolution and line edge roughness (LER) during the subsequent baking process (PEB). Therefore, accurately obtaining the pKa value of the acid generated by PAG is of great significance for photoresist material design.

[0003] In existing technologies, the methods for obtaining the pKa value of photoacid-producing agents mainly rely on experimental titration measurements or quantum chemical calculations based on density functional theory (DFT). However, experimental measurements are costly and have long synthesis cycles, and the pKa values ​​of many superacids used in high-end photolithography (such as perfluoroalkyl sulfonic acids) are difficult to measure accurately in conventional solvents. While molecular dynamics and DFT calculations can provide theoretical guidance, their calculation processes involve complex solvation models and conformational searches, consuming a large amount of computational resources and time, which cannot meet the needs of high-throughput screening.

[0004] Besides experimental and computational costs, data scarcity and molecular complexity further restrict the application of conventional deep learning methods. Experimental determination of the pKa values ​​of photoresist PAG compounds is costly and time-consuming, with typically fewer than a few hundred data points available for modeling, representing a typical small-sample regression scenario. Meanwhile, PAG molecules often contain fluoroalkyl, sulfonamide, and phenolic hydroxyl structural units, and their acidity is jointly governed by the inductive effects of local functional groups and the conjugation transfer of the molecular skeleton. While a single graph neural network can encode the adjacency relationships of atomic nodes, it struggles to perceive the electronic effects of distant substituents at the molecular level; relying solely on global fingerprints easily leads to the loss of crucial local substructural information. Existing single characterization architectures often suffer from amplified prediction biases when dealing with such complex fluorinated molecules due to incomplete feature capture. Summary of the Invention

[0005] The purpose of this application is to provide a machine learning-based method for predicting the dissociation constant of photoacid-producing agents, in order to solve the technical problems in the prior art, such as the high computational resources and time required to calculate pKa values ​​based on molecular dynamics or density functional theory (DFT), the high cost of experimental measurement and the limitation by solvents, and the incomplete feature capture and insufficient prediction accuracy of existing deep learning methods that use a single molecular characterization architecture.

[0006] To achieve the above objectives, the technical solution adopted in this application is as follows: In a first aspect, a machine learning-based method for predicting the dissociation constant of photoacid-producing agents is provided, comprising: obtaining the SMILES formula of a target photoacid-producing agent or the acid-producing molecule of the target photoacid-producing agent; converting the SMILES formula into a two-dimensional descriptor of the molecule based on molecular informatics tools, the two-dimensional descriptor including molecular graph features and a molecular global structure fingerprint, wherein the molecular graph features are used to capture local topological interaction information at the neighborhood level of atomic nodes, and the molecular global structure fingerprint is used to encode the existence information of substructure fragments of the target photoacid-producing agent or the acid-producing molecule as a whole; and inputting the molecular graph features and the molecular global structure fingerprint as input data into a pre-trained... In the constructed dual-branch deep learning model, the dual-branch deep learning model outputs the predicted value of the dissociation constant of the target photoacid-producing agent; wherein, the dual-branch deep learning model includes a graph neural network branch, a multilayer perceptron branch, a feature fusion layer, and a fully connected layer. The graph neural network branch and the multilayer perceptron branch independently encode the molecular graph features and the molecular global structure fingerprint, respectively. The feature fusion layer concatenates the encoding results of the graph neural network branch and the multilayer perceptron branch along the feature dimension to obtain a fused feature vector. The fully connected layer performs dimensionality reduction processing on the fused feature vector and outputs the predicted value of the dissociation constant of the target photoacid-producing agent.

[0007] In one embodiment of this application, the dual-branch deep learning model is trained using a transfer learning strategy, which includes a pre-training stage and a fine-tuning stage. In the pre-training stage, a pre-training data is used to train an initial deep learning model to obtain a pre-trained model. In the fine-tuning stage, fine-tuning data with a smaller amount of data than the pre-training data is used to continue training the pre-trained model. The training process reuses the target value standardization parameters obtained in the pre-training stage to standardize the output space, resulting in a candidate deep learning model. The dual-branch deep learning model is the candidate deep learning model with the highest performance evaluation index among multiple candidate deep learning models. These multiple candidate deep learning models are obtained by independently running the transfer learning strategy on a fixed training set, validation set, and test set using at least three different random seeds. The training set, validation set, and test set are obtained by partitioning the pre-training data and the fine-tuning data. The performance evaluation index is the test set determination coefficient R².

[0008] In one embodiment of this application, the learning rate in the pre-training stage is a fixed learning rate, the batch size is a first preset batch value, and the early stopping tolerance is set to a first preset number of rounds; the learning rate in the fine-tuning stage linearly increases from an initial value to a peak learning rate within a preset number of warm-up rounds, and then decays along a cosine curve from the peak learning rate after reaching the preset number of warm-up rounds, the batch size is a second preset batch value, the early stopping tolerance is set to a second preset number of rounds, the peak learning rate is lower than the fixed learning rate, the second preset batch value is less than the first preset batch value, and the second preset number of rounds is greater than the first preset number of rounds.

[0009] In one embodiment of this application, the pre-training data is obtained by screening from an open-source experimental database, including two-dimensional descriptors and dissociation constants of a number of screened acidic small molecules. The screened acidic small molecules are all molecules whose molecular fingerprint similarity to the target photoacid generator is within a preset range. The fine-tuning data is obtained by experimental titration, including two-dimensional descriptors and dissociation constants of a number of nonionic photoacid generators, and two-dimensional descriptors and dissociation constants of acid molecules corresponding to the anionic portions of a number of ionic photoacid generators. The number of nonionic photoacid generators includes sulfonate acid generators, and the number of acid molecules corresponding to the anionic portions of the number of ionic photoacid generators includes fluoroalkyl sulfonic acids, aryl sulfonic acids, carboxylic acids, and carbanion acid molecules.

[0010] In one embodiment of this application, the molecular fingerprint similarity is the Tanimoto similarity calculated based on Morgan's molecular fingerprint, and the preset range is 0.2 to 0.8; the open-source experimental database is one or more of the databases ChEMBL, LigpKa, SPARC, DataWarrior, iBonD, PubChem, PKADB, DrugBank, and BindingDB.

[0011] In one embodiment of this application, converting the SMILES formula into a two-dimensional descriptor of a molecule based on molecular informatics tools includes: parsing the SMILES formula into a graph object composed of atomic nodes and chemical bond edges based on molecular informatics tools; extracting the physicochemical properties of the atomic nodes, the physicochemical properties of the chemical bonds, and the connection relationships of the chemical bonds; encoding the physicochemical properties of the atomic nodes, the physicochemical properties of the chemical bonds, and the connection relationships of the chemical bonds respectively to form a node feature matrix, an edge feature matrix, and an edge index matrix describing the connection topology between atomic nodes; and constructing the molecular graph feature by the node feature matrix, the edge feature matrix, and the edge index matrix; and encoding the topological environment or substructure fragments of the atomic node neighborhood of the molecule corresponding to the SMILES formula based on the molecular informatics tools, and mapping the encoding result into a fixed-length vector through a hash function to obtain the global structural fingerprint of the molecule.

[0012] In one embodiment of this application, the molecular global structure fingerprint includes one or more of the following: molecular informatics tool fingerprint, Morgan molecular fingerprint, MACCS fingerprint, atom pair fingerprint, topological torsion fingerprint, and pattern fingerprint.

[0013] In one embodiment of this application, the graph neural network branch employs one or more of the following: attention-based graph neural networks, equivariant graph neural networks, continuous filter convolutional networks, and graph isomorphic networks.

[0014] The beneficial effects of this invention include: First, it efficiently predicts dissociation constants to meet high-throughput screening requirements: By replacing existing density functional theory quantum chemical calculations or experimental titration methods with a dual-branch deep learning model, the predicted value of the target dissociation constant can be obtained quickly by simply inputting the SMILES formula of the target photoacid generator. This significantly shortens the time required to obtain the pKa value of a specific photoacid generator, and is especially suitable for high-throughput evaluation and screening of large-scale virtual molecular libraries for monolayer positive photoresist systems, significantly reducing R&D costs and accelerating the iteration cycle.

[0015] Second, the dual-branch feature fusion provides superior prediction accuracy compared to a single representation architecture: A dual-branch deep learning model is employed, where molecular graph features and global molecular structural fingerprints are independently encoded via graph neural network (GNN) and multilayer perceptron (MLP) branches, respectively. A feature fusion layer then concatenates the encoding results from the GNN and MLP branches along the feature dimension to obtain a fused feature vector. Finally, a fully connected layer performs dimensionality reduction on the fused feature vector to output the predicted dissociation constant. This dual-branch deep learning model simultaneously possesses sensitivity to local electronic effects of functional groups and global perception capabilities of the overall conjugated molecular system. This architecture can explain the structure-activity relationship of complex, large-volume photoacid-producing agents from a multi-dimensional "local + global" perspective, solving the problem of incomplete capture of local topological and global structural information and insufficient prediction accuracy in single representation architectures.

[0016] Third, overcome the limitations of small samples and improve generalization ability: By introducing a transfer learning strategy into the deep learning architecture, pre-training is performed using broad-spectrum chemical knowledge from open-source experimental databases, and then fine-tuning is performed on a small dataset of photoacid generators. At the same time, in the fine-tuning stage, the target values ​​obtained in the pre-training stage are reused to standardize parameters to eliminate the output space distribution offset. Combined with a learning rate scheduling strategy of linear preheating combined with cosine decay, the model can fully explore the parameter space under small sample conditions without skipping the optimal solution region due to excessive step size. This effectively solves the technical pain point in the development of advanced photoresists where the model is prone to overfitting due to the scarcity of accurate pKa values ​​of superacids in photoacid generators. Attached Figure Description

[0017] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of a machine learning-based method for predicting the dissociation constant of photoacid-producing agents according to an embodiment of this application; Figure 2 This is a graph showing the fit between the predicted and experimental values ​​of the AttentiveFP graph neural network model on a real dataset. Figure 3 This is a graph showing the fitting of predicted values ​​and experimental values ​​of a multilayer perceptron (MLP) on a real dataset. Figure 4 This is a graph showing the fitting of the predicted values ​​and experimental values ​​of the dual-branch deep learning model provided by this solution on a real dataset; Figure 5 This is an example of an interpretability analysis plot of the integral gradient (IGs) of a molecule provided in this application; Figure 6 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein, and therefore this application is not limited to the specific embodiments disclosed below.

[0020] As illustrated in this application, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0021] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously. Furthermore, other operations may be added to these processes, or one or more steps may be removed from these processes.

[0022] Figure 1 This is a flowchart of a machine learning-based method for predicting the dissociation constant of photoacid-producing agents according to an embodiment of this application. See also... Figure 1 This includes the following steps: Step S110: Obtain the SMILES formula of the target photoacid-producing agent or the acid-producing molecule of the target photoacid-producing agent.

[0023] Among them, the target photoacid-generating agent refers to the photoacid-generating compound used as a photosensitive component in photoresist. SMILES, or Molecular Linear Input Specification, is a specification that uses ASCII strings to describe molecular structures. It linearly represents the two-dimensional topological structure of molecules through atomic node symbols, chemical bond symbols, and branch brackets, achieving lossless digital representation of molecular information and providing standardized underlying input for subsequent machine learning.

[0024] Step S120: Based on molecular informatics tools, the SMILES formula is converted into a two-dimensional descriptor of the molecule.

[0025] The two-dimensional descriptor includes molecular graph features and molecular global structure fingerprints.

[0026] In one embodiment of this application, converting the SMILES formula into a two-dimensional descriptor of a molecule based on molecular informatics tools includes: parsing the SMILES formula into a graph object composed of atomic nodes and chemical bond edges based on molecular informatics tools; extracting the physicochemical properties of atomic nodes and the physicochemical properties and connection relationships of chemical bonds; encoding the physicochemical properties of atomic nodes, the physicochemical properties and connection relationships of chemical bonds respectively to form a node feature matrix, an edge feature matrix, and an edge index matrix describing the connection topology between atomic nodes; and constructing molecular graph features from the node feature matrix, the edge feature matrix, and the edge index matrix; and encoding the topological environment or substructure fragments of the atomic node neighborhood of the molecule corresponding to the SMILES formula based on molecular informatics tools, and mapping the encoding result into a fixed-length vector through a hash function to obtain the global structural fingerprint of the molecule.

[0027] Specifically, molecular informatics tools refer to software toolkits used for cheminformatics processing, such as the RDKit toolkit. The image object is... In this representation, A represents an atom and E represents a chemical bond. This method clearly depicts the spatial relationships between atoms and the connections between chemical bonds, effectively capturing the topological structure of molecules. The physicochemical properties of atomic nodes include atom type, atomic number, formal charge, hybridization mode, and whether it is in an aromatic ring, etc.; the physicochemical properties of chemical bonds include bond type (single, double, triple, aromatic), and whether it is in a conjugated system, etc. The connections between chemical bonds are represented by an edge index matrix, recording the indices of the two atoms connected by each edge. This method enables the systematic extraction of molecular graph features to capture local topological interaction information at the neighborhood level of atomic nodes.

[0028] Molecular global structural fingerprints can be a mixture of different fingerprints. In some embodiments, molecular global structural fingerprints include one or more of the following: molecular informatics tool fingerprints (i.e., RDKit fingerprints), Morgan molecular fingerprints, MACCS fingerprints, atom-pair fingerprints, topological torsion fingerprints, and pattern fingerprints. For example, a combination of a 512-bit Morgan molecular fingerprint and a 466-dimensional molecular informatics tool fingerprint can be used to provide a more comprehensive expression of molecular physicochemical features. This approach can effectively extract molecular global structural fingerprints to encode the presence of substructural fragments of the target photoacid-producing agent or the acid-producing molecule as a whole. Molecular graph features and molecular global structural fingerprints form complementary characterizations, jointly improving the accuracy and generalization ability of acid pKa value prediction.

[0029] Step S130: The molecular graph features and the global molecular structure fingerprint are used as input data and input into the trained dual-branch deep learning model, so that the dual-branch deep learning model outputs the predicted value of the dissociation constant of the target photoacid-producing agent.

[0030] The dual-branch deep learning model includes a graph neural network (AttentiveFP) branch, a multilayer perceptron (MLP) branch, a feature fusion layer, and a fully connected layer. The graph neural network branch and the multilayer perceptron branch independently encode the molecular graph features and the molecular global structure fingerprint, respectively. The feature fusion layer concatenates the encoding results of the graph neural network branch and the multilayer perceptron branch along the feature dimension to obtain a fused feature vector. The fully connected layer performs dimensionality reduction on the fused feature vector and outputs the predicted value of the dissociation constant of the target photoacid-producing agent.

[0031] In one embodiment of this application, the graph neural network branch employs one or more of the following: attention-based graph neural networks, equivariant graph neural networks, continuous filter convolutional networks, and graph isomorphic networks.

[0032] The dual-branch deep learning model achieves higher prediction accuracy compared to existing deep learning models employing a single-molecule representation architecture. Its underlying principle is as follows: Graph neural networks excel at capturing local topological interactions between atoms and the inductive effects of functional groups in molecular diagrams, but have limited ability to perceive the electronic effects of distant substituents; multilayer perceptrons, by encoding the existence of substructural fragments of the entire molecule through global fingerprinting, can compensate for the shortcomings of graph neural networks in global structure perception. The two feature paths, after independent encoding, converge in a fusion layer, enabling the model to simultaneously possess sensitivity to local electronic effects of functional groups and global perception of the conjugated system of the entire molecule. This allows for a multi-dimensional "local + global" explanation of the structure-activity relationship of complex, large-volume photoacid-producing agents.

[0033] The training method of the dual-branch deep learning model is explained in detail below with reference to the accompanying diagram.

[0034] In one embodiment of this application, the dual-branch deep learning model is trained using a transfer learning strategy, which includes a pre-training phase and a fine-tuning phase. In the pre-training phase, the initial deep learning model is trained using pre-training data to obtain a pre-trained model. In the fine-tuning phase, the pre-trained model is trained again using fine-tuning data with a smaller amount of data than the pre-training data. The training process reuses the target value standardization parameters obtained in the pre-training phase to standardize the output space, resulting in a candidate deep learning model. The dual-branch deep learning model is the candidate deep learning model with the highest performance evaluation index among multiple candidate deep learning models. Multiple candidate deep learning models are obtained by independently running the transfer learning strategy on a fixed training set, validation set, and test set with no less than three different random seeds. The training set, validation set, and test set are obtained by dividing the data using pre-training data and fine-tuning data. The performance evaluation index is the test set determination coefficient R².

[0035] For example, both pre-training and fine-tuning data are fixedly divided in an 8:1:1 ratio, with the training set accounting for 80%, the test set for 10%, and the validation set for 10%. During the pre-training phase, the initial deep learning model is trained using the training set of the pre-training data; during the fine-tuning phase, the pre-trained model is further trained using the training set of the fine-tuning data. Performance evaluation of each candidate deep learning model is performed on the test set, ensuring that the data from the training, validation, and evaluation phases are independent of each other.

[0036] The advantages of this approach are twofold: First, by learning the dissociation rules of acidic compounds on a large-scale, broad-spectrum acidic molecule dataset during the pre-training phase, the model acquires general knowledge of structure-acidity relationships. Then, during the fine-tuning phase, the pre-trained model is adapted using a high-precision, small dataset of photoacid-producing agents, effectively overcoming the problem of overfitting in small-sample scenarios caused by the scarcity of pKa values ​​for photoacid-producing agents, and improving the model's generalization ability in small-sample scenarios. Second, the fine-tuning phase reuses the standardized parameters of the target values ​​obtained in the pre-training phase to standardize the output space, eliminating the output space distribution offset caused by different data distributions between the pre-training and fine-tuning phases. This eliminates the need for refitting the standardized parameters during fine-tuning, ensuring that the model can perform its functions effectively. The consistency of the output space reduces the instability of parameter estimation under small sample conditions. Third, the complete training process is run independently on a fixed dataset partition with no less than three different random seeds, eliminating the randomness of training results caused by a single random seed and ensuring the statistical reliability of model performance evaluation. In the small sample scenario of predicting the pKa value of photoacid-producing agents, the impact of random seeds on training results is particularly significant. The strategy of repeatedly selecting the optimal model independently multiple times effectively reduces the risk of model selection. Fourth, the fixed division of training, validation and test sets ensures that the only variable between multiple runs is the random seed, making the performance evaluation of each candidate deep learning model comparable on the same test set, and the selection result is objective and reliable.

[0037] The pre-training and fine-tuning phases can share the same early stopping mechanism. When the decrease in the root mean square error of the validation set is less than the preset decrease threshold or shows an upward trend in consecutive tolerance rounds, the current phase of training is terminated, and the model weights are rolled back to the lowest historical value of the root mean square error of the validation set, in order to prevent the model from overfitting and retain the model parameters with the best performance on the validation set.

[0038] In one embodiment of this application, the learning rate in the pre-training stage is a fixed learning rate, the batch size is a first preset batch value, and the early stopping tolerance is set to a first preset number of rounds; in the fine-tuning stage, the learning rate linearly increases from the initial value to the peak learning rate within a preset number of warm-up rounds, and then decays along a cosine curve from the peak learning rate after reaching the preset number of warm-up rounds, the batch size is a second preset batch value, the early stopping tolerance is set to a second preset number of rounds, the peak learning rate is lower than the fixed learning rate, the second preset batch value is less than the first preset batch value, and the second preset number of rounds is greater than the first preset number of rounds.

[0039] For example, the fixed learning rate is 1×10 -4 The first preset batch size is 512, the first preset number of rounds is 5, and the peak learning rate is 5×10. -5 The preset preheating cycle number is 10, the second preset batch size is 32, and the second preset cycle number is 30.

[0040] The advantage of this approach is that, during the pre-training phase, the fixed learning rate allows the model to be trained sufficiently on large-scale pre-training data with a stable step size, avoiding oscillations in the training process caused by learning rate fluctuations. At the same time, a larger batch size helps to accelerate the convergence speed during the pre-training phase. During the fine-tuning phase, the learning rate is reduced to below the peak learning rate during the pre-training phase, and the batch size is reduced, allowing the model to update parameters with a finer step size on small sample fine-tuning data, thereby improving the generalization ability under small sample conditions. Furthermore, during the fine-tuning phase, the learning rate linearly increases from its initial value to its peak learning rate within a preset number of warm-up rounds. After reaching the preset number of warm-up rounds, it decays from the peak learning rate along a cosine curve. At the same time, the early stopping tolerance is relaxed from the first preset number of rounds in the pre-training phase to the second preset number of rounds. This allows the model to avoid drastic parameter fluctuations in the early stages of training through linear warm-up, and to finely converge with gradually decreasing step sizes through cosine decay in the later stages of training. The relaxed early stopping tolerance also adapts to the characteristics of large fluctuations in validation set metrics in small sample scenarios, avoiding premature termination of training due to short-term metric fluctuations. This prevents overfitting while preserving the model parameters with the best performance on the validation set.

[0041] In one embodiment of this application, the pre-training data is obtained by screening from an open-source experimental database, including two-dimensional descriptors and dissociation constants of a number of screened acidic small molecules. The screened acidic small molecules are all molecules whose molecular fingerprint similarity to the target photoacid generator is within a preset range. The fine-tuning data is obtained by experimental titration, including two-dimensional descriptors and dissociation constants of a number of nonionic photoacid generators, and two-dimensional descriptors and dissociation constants of acid molecules corresponding to the anionic portions of a number of ionic photoacid generators. The number of nonionic photoacid generators includes sulfonate acid generators, and the number of acid molecules corresponding to the anionic portions of a number of ionic photoacid generators includes fluoroalkyl sulfonic acids, aryl sulfonic acids, carboxylic acids, and carbanion acid molecules.

[0042] Specifically, the two-dimensional descriptors for nonionic photoacid-producing agents in the fine-tuning data are obtained by converting the SMILES formulas of the nonionic photoacid-producing agents themselves using molecular informatics tools. Ionic photoacid-producing agents in the fine-tuning data consist of cationic and anionic moieties. The cationic moieties have no substantial impact on the pKa value; the pKa value is determined by the conjugate acid of the anionic moieties. Therefore, the SMILES formulas of the acid molecules corresponding to the anionic moieties are extracted and converted into two-dimensional descriptors for model training. This approach eliminates the interference of irrelevant structural information on model predictions, helping to improve the accuracy of the prediction results.

[0043] In one embodiment of this application, the molecular fingerprint similarity is the Tanimoto similarity calculated based on Morgan's molecular fingerprint, with a preset range of 0.2 to 0.8. The open-source experimental database is one or more of the following databases: ChEMBL, LigpKa, SPARC, DataWarrior, iBonD, PubChem, PKADB, DrugBank, and BindingDB.

[0044] For example, the pre-training data comes from open-source experimental databases such as ChEMBL. The RDKit toolkit is used to calculate the Tanimoto similarity between the Morgan molecular fingerprint of the target photoacid-producing agent and the Morgan molecular fingerprints of various acidic small molecules in the open-source experimental database. Acidic small molecules with Tanimoto similarity ranging from 0.2 to 0.8 (11,566 molecules in total) are selected, and their two-dimensional descriptors and pKa values ​​are used as pre-training data. The fine-tuning data comes from experimental titrations. The two-dimensional descriptors and pKa values ​​of several specific non-ionic photoacid-producing agents, and the two-dimensional descriptors and pKa values ​​of the acid molecules corresponding to the anionic portions of several specific ionic photoacid-producing agents (such as perfluoroalkyl sulfonic acid polycyclic or fluorinated compounds, 229 molecules in total) are used as fine-tuning data.

[0045] Furthermore, before model training, SMILES-style one-dimensional descriptors can be used to examine the reasonableness of the dataset distribution for pre-training and fine-tuning data. A one-dimensional descriptor is a descriptor that represents the overall physicochemical properties of a molecule with a single numerical value, including one or more of molecular weight, partition coefficient, number of hydrogen bond acceptors, and number of hydrogen bond donors. If the dataset distribution is reasonable, the pre-training and fine-tuning data can be used for subsequent training of the two-branch deep learning model; if the dataset distribution is too disparate, the preset screening range of molecular fingerprint similarity can be adjusted before further examination.

[0046] The advantage of this approach is that the pre-training data is screened using molecular fingerprint similarity, ensuring that the acidic small molecules in the pre-training dataset are structurally similar to the target photoacid-producing agent. This allows the structure-acidity relationship learned during pre-training to be effectively transferred to the pKa prediction task of photoacid-producing agents, avoiding negative transfer problems caused by excessive differences in the chemical space between the pre-training and fine-tuning data. Simultaneously, the fine-tuning data covers both non-ionic and ionic photoacid-producing agents, enabling the model to predict various types of photoacid-producing agents.

[0047] To verify the prediction accuracy optimization effect of the dual-branch deep learning model provided in this solution, two single-branch models (AttentiveFP and MLP graph neural network models) were tested on the same real dataset along with the dual-branch deep learning model provided in this solution. Specifically, Figure 2 This is a graph showing the fit between the predicted and experimental values ​​of the AttentiveFP graph neural network model on a real dataset. Figure 3 This is a graph showing the fit between the predictions and experimental values ​​of a multilayer perceptron (MLP) on a real dataset. Figure 4 This is a graph showing the fitting of the predicted values ​​and experimental values ​​of the dual-branch deep learning model provided in this solution on a real dataset. See also... Figures 2 to 4 The x-axis represents the experimentally determined pKa value, the y-axis represents the model-predicted pKa value, the cyan scatter plots represent the training set data, the blue scatter plots represent the test set data, and the dashed line represents the ideal fit line. R0 2 Here, is the coefficient of determination for the test set, MAE is the mean absolute error, and RMSE is the root mean square error. It can be seen that... Figure 4 The scattered points are most closely distributed near the ideal fitting line, and the root mean square error and mean absolute error of the two-branch deep learning model are significantly lower than those of the two single-branch models, indicating that the prediction accuracy of the two-branch deep learning model is significantly better than that of any single-branch model, thus verifying the effectiveness of the two-branch architecture.

[0048] Furthermore, to verify the authenticity of the molecular structure-acidity relationship learned by the bi-branch deep learning model, a Y-randomization test can be performed. The specific process of the Y-randomization test is as follows: the pKa labels in the training set, validation set, and test set are merged and then randomly shuffled. The sets are then re-divided according to their original sizes, while maintaining the molecular graph features and global molecular structure fingerprint. The model is then retrained and evaluated with the same training hyperparameters. This process is repeated at least 25 times independently. If the coefficient of determination R² of the retrained model on the test set drops to zero or a negative value, it proves that the bi-branch deep learning model has learned a real molecular structure-acidity relationship rather than a random association.

[0049] Furthermore, in order to provide quantitative guidance for the reverse design of photoacid-producing agent molecules, the bi-branch deep learning model can also perform molecular-level interpretability attribution analysis of the prediction results using the integral gradient (IGs) method. Figure 5 This is an example of an interpretability analysis plot of the integral gradient (IGs) of a molecule provided in this application. Figure 5As shown, the dual-branch deep learning model accurately labels the contribution of fluorine atoms to the predicted pKa value as negative through integral gradient calculation, meaning the presence of fluorine atoms reduces the predicted pKa value. Furthermore, by performing prediction and integral gradient attribution separately for photoacid-producing molecules with fluorine atoms at different positions (e.g., varying distances from the dissociation center), it can be observed that the predicted pKa value and the negative contribution of the integral gradient to fluorine atoms dynamically change with the distance between the fluorine atom and the dissociation center. These observations are consistent with chemical theory, verifying the reliability of the attribution results. Based on the direction and magnitude of each atom's contribution, the structural modification strategy can be determined in reverse, providing an intuitive and quantitative basis for precise control of the photoresist acid diffusion coefficient through molecular reverse design.

[0050] Figure 6 This is a schematic diagram of an electronic device provided according to an embodiment of this application. (Reference) Figure 6 As shown, the electronic device 600 may include an internal communication bus 601, a processor 602, a read-only memory (ROM) 603, a random access memory (RAM) 604, and a communication port 605. When applied to a personal computer, the electronic device 600 may also include a hard disk 606. The internal communication bus 601 enables data communication between components of the electronic device 600. The processor 602 can make judgments and issue prompts. In some embodiments, the processor 602 may consist of one or more processors. The communication port 605 enables data communication between the electronic device 600 and external devices. In some embodiments, the electronic device 600 can send and receive information and data from a network through the communication port 605. The electronic device 600 may also include different forms of program storage units and data storage units, such as the hard disk 606, the read-only memory (ROM) 603, and the random access memory (RAM) 604, capable of storing various data files used for computer processing and / or communication, as well as possible program instructions executed by the processor 602. The processor executes these instructions to implement the main part of the method. The results of the processor processing are transmitted to the user device through the communication port and displayed on the user interface.

[0051] Figure 1 The method for predicting the dissociation constant of the photoacid-producing agent shown can be implemented as a computer program, stored in hard disk 606, and loaded into processor 602 for execution.

[0052] This application also includes a computer-readable medium storing computer program code that, when executed by a processor, implements the aforementioned method for predicting the dissociation constant of photoacid-producing agents.

[0053] When the method for predicting the dissociation constant of photoacid-producing agents is implemented as a computer program, it can also be stored as an article of manufacture in a computer-readable storage medium. For example, computer-readable storage media can include, but are not limited to, magnetic storage devices (e.g., hard disks, floppy disks, magnetic stripes), optical discs (e.g., compact discs (CDs), digital multifunction discs (DVDs)), smart cards, and flash memory devices (e.g., electrically erasable programmable read-only memory (EPROM), cards, sticks, key drives). Furthermore, the various storage media described herein can represent one or more devices and / or other machine-readable media used for storing information. The term "machine-readable medium" can include, but is not limited to, wireless channels and various other media (and / or storage media) capable of storing, containing, and / or carrying code and / or instructions and / or data.

[0054] It should be understood that the embodiments described above are merely illustrative. The embodiments described herein may be implemented in hardware, software, firmware, middleware, microcode, or any combination thereof. For hardware implementation, the processor may be implemented within one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, and / or other electronic units designed to perform the functions described herein, or combinations thereof.

[0055] Some aspects of this application can be executed entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The aforementioned hardware or software may be referred to as a "data block," "module," "engine," "unit," "component," or "system." The processor may be one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DAPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, or combinations thereof. Furthermore, aspects of this application may manifest as computer products residing in one or more computer-readable media, including computer-readable program code. For example, computer-readable media may include, but are not limited to, magnetic storage devices (e.g., hard disks, floppy disks, magnetic tapes, etc.), optical discs (e.g., compressed CDs, digital multifunction DVDs, etc.), smart cards, and flash memory devices (e.g., cards, sticks, key drives, etc.).

[0056] A computer-readable medium may contain a propagated data signal containing computer program code, for example, on baseband or as part of a carrier wave. This propagated signal may take various forms, including electromagnetic, optical, and so on, or suitable combinations thereof. A computer-readable medium can be any computer-readable medium other than a computer-readable storage medium, which can be connected to an instruction execution system, apparatus, or device to enable communication, propagation, or transmission of a program for use. The program code located on the computer-readable medium can be propagated through any suitable medium, including radio, cable, fiber optic cable, radio frequency signals, or similar media, or any combination of the above media.

[0057] The basic concepts have been described above. Obviously, for those skilled in the art, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore remain within the spirit and scope of the exemplary embodiments of this application.

[0058] Furthermore, this application uses specific terms to describe embodiments of the application. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of the application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different locations in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of the application can be appropriately combined.

[0059] In some embodiments, numbers describing the quantity of components and attributes are used. It should be understood that such numbers used to describe embodiments are sometimes modified by the terms "approximately," "approximately," or "generally." Unless otherwise stated, "approximately," "approximately," or "generally" indicates that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in this application are approximate values, which may be changed depending on the characteristics required by individual embodiments. In some embodiments, numerical parameters should take into account specified significant digits and employ a general method of digit reservation. Although the numerical ranges and parameters used to confirm their breadth of range in some embodiments of this application are approximate values, in specific embodiments, such values ​​are set as precisely as feasible.

Claims

1. A method for predicting the dissociation constant of photoacid-producing agents based on machine learning, characterized in that, include: Obtain the SMILES formula of the target photoacid-producing agent or the acid-producing molecule of the target photoacid-producing agent; The SMILES formula is transformed into a two-dimensional descriptor of the molecule using molecular informatics tools. The two-dimensional descriptor includes molecular graph features and molecular global structure fingerprint. The molecular graph features are used to capture local topological interaction information at the neighborhood level of atomic nodes. The molecular global structure fingerprint is used to encode the existence information of substructure fragments of the target photoacid-producing agent or the acid-producing molecule as a whole. The molecular graph features and the molecular global structure fingerprint are used as input data and input into the trained bi-branch deep learning model, so that the bi-branch deep learning model outputs the predicted value of the dissociation constant of the target photoacid-producing agent. The dual-branch deep learning model includes a graph neural network branch, a multilayer perceptron branch, a feature fusion layer, and a fully connected layer. The graph neural network branch and the multilayer perceptron branch independently encode the molecular graph features and the molecular global structure fingerprint, respectively. The feature fusion layer concatenates the encoding results of the graph neural network branch and the multilayer perceptron branch along the feature dimension to obtain a fused feature vector. The fully connected layer performs dimensionality reduction processing on the fused feature vector and outputs the predicted value of the dissociation constant of the target photoacid-producing agent.

2. The method for predicting the dissociation constant of photoacid-producing agents based on machine learning according to claim 1, characterized in that, The dual-branch deep learning model is trained using a transfer learning strategy, which includes a pre-training phase and a fine-tuning phase. In the pre-training phase, the initial deep learning model is trained using pre-training data to obtain a pre-trained model; In the fine-tuning stage, the pre-trained model is trained again using fine-tuning data with a smaller amount of data than the pre-training data. The training process reuses the target value standardization parameters obtained in the pre-training stage to standardize the output space, thereby obtaining a candidate deep learning model. The dual-branch deep learning model is the candidate deep learning model with the highest performance evaluation index among the multiple candidate deep learning models. The multiple candidate deep learning models are obtained by independently running the transfer learning strategy on a fixed training set, validation set, and test set with no less than three different random seeds. The training set, validation set, and test set are obtained by dividing the pre-training data and the fine-tuning data. The performance evaluation index is the test set determination coefficient R².

3. The method for predicting the dissociation constant of photoacid-producing agents based on machine learning according to claim 2, characterized in that, The learning rate in the pre-training phase is a fixed learning rate, the batch size is a first preset batch value, and the early stopping tolerance is set to a first preset number of rounds. The learning rate in the fine-tuning phase linearly increases from the initial value to the peak learning rate within a preset number of warm-up rounds, and then decays along a cosine curve from the peak learning rate after reaching the preset number of warm-up rounds. The batch size is a second preset batch value, the early stop tolerance is set to a second preset number of rounds, the peak learning rate is lower than the fixed learning rate, the second preset batch value is less than the first preset batch value, and the second preset number of rounds is greater than the first preset number of rounds.

4. The method for predicting the dissociation constant of photoacid-producing agents based on machine learning according to claim 2, characterized in that, The pre-training data was obtained by screening from an open-source experimental database, including two-dimensional descriptors and dissociation constants of a number of screened acidic small molecules. The screened acidic small molecules are all molecules whose molecular fingerprint similarity to the target photoacid generator is within a preset range. The fine-tuning data was obtained through experimental titration and included two-dimensional descriptors and dissociation constants of several nonionic photoacid-producing agents, and two-dimensional descriptors and dissociation constants of acid molecules corresponding to the anionic portions of several ionic photoacid-producing agents. The several nonionic photoacid-producing agents included sulfonate acid-producing agents, and the acid molecules corresponding to the anionic portions of the several ionic photoacid-producing agents included fluoroalkyl sulfonic acids, aryl sulfonic acids, carboxylic acids, and carbanion acid molecules.

5. The method for predicting the dissociation constant of photoacid-producing agents based on machine learning according to claim 4, characterized in that, The molecular fingerprint similarity is the Tanimoto similarity calculated based on Morgan's molecular fingerprint, and the preset range is 0.2 to 0.8; The open-source experimental database is one or more of the following databases: ChEMBL, LigpKa, SPARC, DataWarrior, iBonD, PubChem, PKADB, DrugBank, and BindingDB.

6. The method for predicting the dissociation constant of photoacid-producing agents based on machine learning according to claim 1, characterized in that, The conversion of the SMILES formula into a two-dimensional molecular descriptor based on molecular informatics tools includes: Based on molecular informatics tools, the SMILES formula is parsed into a graph object composed of atomic nodes and chemical bond edges. The physicochemical properties of the atomic nodes, the physicochemical properties of the chemical bonds, and the connection relationships of the chemical bonds are extracted. These properties are then encoded to form a node feature matrix, an edge feature matrix, and an edge index matrix describing the topology of connections between atomic nodes. The molecular graph features are constructed from these matrixes. Based on the molecular informatics tool, the topological environment or substructure fragments of the atomic node neighborhood of the SMILES-type molecule are encoded, and the encoding result is mapped into a fixed-length vector through a hash function to obtain the global structural fingerprint of the molecule.

7. The method for predicting the dissociation constant of photoacid-producing agents based on machine learning according to claim 1, characterized in that, The molecular global structure fingerprint includes one or more of the following: molecular informatics tool fingerprint, Morgan molecular fingerprint, MACCS fingerprint, atom pair fingerprint, topological torsion fingerprint, and pattern fingerprint.

8. The method for predicting the dissociation constant of photoacid-producing agents based on machine learning according to claim 1, characterized in that, The graph neural network branch employs one or more of the following: attention-based graph neural networks, equivariant graph neural networks, continuous filter convolutional networks, and graph isomorphic networks.

9. An electronic device, characterized in that, include: One or more processors; as well as, One or more memories coupled to one or more processors and storing instructions thereon, which, when the instructions are executed individually or jointly by one or more processors, cause the electronic device to perform the method as described in any one of claims 1-8.

10. A non-transitory computer-readable storage medium storing machine-executable instructions, characterized in that, When executed by one or more processors of the machine, the machine-executable instructions cause the machine to perform the method as described in any one of claims 1-8.