Physical Constraint Frameworks in Chemistry: From Traditional Descriptors to Equivariant Neural Networks
Communications in Computational Chemistry · 09.05.2026 · 8 (3) · 249-273
DOI: 10.4208/cicc.2026.83.01 ↗Kurz gefasst
Fachartikel in der Zeitschrift Communications in Computational Chemistry, online erschienen am 09.05.2026. Verfasst von Zhenxing Zhu und 4 weiteren Autor:innen. Die Arbeit nennt 106 Literaturangaben.
Abstract (Originalsprache)
Machine learning has become an indispensable component of modern computational chemistry and materials science, enabling accurate prediction of molecular energetics, reaction pathways, and solid-state properties. A central challenge in these endeavors is incorporating fundamental physical constraints into data-driven models to ensure reliability and generalizability. Among these constraints, geometric symmetries impose strict requirements: physical models must be invariant under spatial transformations when predicting scalar quantities (e.g., energy), and equivariant when predicting directional observables (e.g., forces or dipoles). This review provides a structured overview of how physical constraints, ranging from soft chemical priors to hard symmetry enforcement, have shaped the evolution of molecular and materials representations from handcrafted descriptors to deep neural networks and, most recently, to symmetry-enforced equivariant architectures. We begin with a rigorous yet accessible treatment of physical constraints, focusing on invariance and equivariance from the perspective of group theory while clarifying their role as stringent geometric requirements. We then survey classical descriptor-based approaches and discuss their continued relevance as effective embedders of chemical intuition. Next, we examine the rise of general-purpose deep learning models (e.g., SchNet, DeePMD) that achieve basic physical consistency through geometric inputs, highlighting their success as robust, widely adopted tools. Finally, we present the theoretical foundations of SE(3)-equivariant networks, emphasizing how spherical harmonics and Clebsch-Gordan couplings enable exact enforcement of geometric physical laws, and review their emerging applications across diverse domains including organic, drug-like molecules, excited-state dynamics, and periodic crystalline materials. Rather than positioning equivariant models as a universal replacement, we argue that they represent a rigorous frontier of physical constraints for tasks demanding high physical fidelity, data efficiency, or tensorial output prediction, while traditional descriptors and invariant deep networks remain effective and dominant in many routine settings where softer constraints suffice. By connecting mathematical principles to practical implementations, this review aims to clarify the relationships, trade-offs, and complementary roles of these methodologies in the ongoing quest for physics-informed machine learning in chemistry and materials science.
Zitieren
APA
Zhu, Z., Zhao, S., Tang, D., Guo, J., & Shen, L. (2026). Physical Constraint Frameworks in Chemistry: From Traditional Descriptors to Equivariant Neural Networks. Communications in Computational Chemistry, 8(3), 249-273. https://doi.org/10.4208/cicc.2026.83.01
Vancouver
Zhu Z, Zhao S, Tang D, Guo J, Shen L. Physical Constraint Frameworks in Chemistry: From Traditional Descriptors to Equivariant Neural Networks. Communications in Computational Chemistry. 2026;8(3):249-273. doi:10.4208/cicc.2026.83.01
BibTeX
@article{Zhu2026,
author = {Zhu, Zhenxing and Zhao, Sirui and Tang, Diandong and Guo, Jingyu and Shen, Lin},
title = {Physical Constraint Frameworks in Chemistry: From Traditional Descriptors to Equivariant Neural Networks},
journal = {Communications in Computational Chemistry},
year = {2026},
volume = {8},
number = {3},
pages = {249--273},
doi = {10.4208/cicc.2026.83.01}
}
EUVIMED