A Kolmogorov-Arnold Surrogate Model for Chemical Equilibria: Application to Solid Solutions

This repository contains supporting material to generate datasets, train neural networks, and analyze their predictions on chemical equilibria. Four chemical systems are examined: Cement_system (i)_mech_mix: pure mechanical mixture (ii)_SS : two-component (Ba,Ra)SO4 solid solution model (iii)_SS_Sr: three-component (Sr,Ba,Ra)SO4 solid solution model The GEMS folder provides the GEM-Selektor input files for each chemical system, as well as the Python scripts that create the datasets and the conda environment. The resulting NumPy arrays are stored in Datasets. Two sample training scripts for KAN and MLP are available in the Training folder. The trained neural network models are saved in Models. Finally, the Analysis folder contains the predictions on the test set associated with each model and two Jupyter Notebooks to generate the error plots.本仓库包含用于生成数据集、训练神经网络以及分析其在化学平衡领域预测结果的辅助材料。本次研究共考察四类化学体系： Cement_system (i)_mech_mix：纯机械混合物 (ii)_SS：二组分(Ba,Ra)SO₄固溶体模型 (iii)_SS_Sr：三组分(Sr,Ba,Ra)SO₄固溶体模型 GEMS 文件夹提供了各化学体系对应的GEM-Selektor输入文件，以及用于生成数据集并配置conda环境的Python脚本。生成的NumPy数组存储于Datasets目录中。Training目录下提供了两份针对KAN与多层感知机（Multi-Layer Perceptron, MLP）的示例训练脚本。训练完成的神经网络模型保存于Models目录。最后，Analysis文件夹包含各模型对应的测试集预测结果，以及两份用于生成误差分布图的Jupyter Notebook文件。