Regularization Strategy for Battery Materials Using Lattice Scaling

Abstract

Artificial iIntelligence (AI) has emerged as a powerful tool for accelerating materials discovery; however, its effectiveness in battery research remains constrained by the limited size and heterogeneity of available materials datasets. In this work, we introduce a physically informed regularization strategy based on lattice scaling of crystalline structures to reduce overfitting and enhance machine-learning performance in lithium-ion battery materials prediction. The proposed framework generates structurally perturbed variants through controlled isotropic and anisotropic modifications of unit-cell volumes within experimentally realistic limits (±5-30%), mimicking chemo-mechanical variations occurring during electrochemical cycling. The method is evaluated using Crystal Graph Convolutional Neural Networks (CGCNN) trained on electrode materials from the Materials Project Battery Explorer database to predict seven electrochemical and four intrinsic material properties. We demonstrate that lattice scaling - particularly anisotropic scaling within the physically meaningful range of ±5% volume variation - systematically improves predictive accuracy compared with both baseline oversampling and established crystal regularization techniques, achieving improvements of up to 10.5% in the MAD/MAE ratio for key electrochemical descriptors. The results highlight the importance of physically meaningful transformations in materials model regularization and reveal a property-dependent response to regularization strategies. The proposed approach provides a simple, computationally efficient, and architecture-independent pathway to improve data efficiency and generalization in AI-driven battery materials discovery.