Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces E3D-IQA, a diagnostic framework that connects the latent edge-energy representation of Allegro-type machine-learning interatomic potentials (MLIPs) with Interacting Quantum Atoms (IQA) energy decomposition. By incorporating intra-atomic supervision, the method reveals that traditional energy and force training alone fails to achieve an IQA-like energy allocation, while the proposed approach successfully aligns node energies with IQA intra-atomic terms and enhances the accuracy of latent edge energies. The findings highlight that residual errors are primarily associated with weak pair interactions, suggesting that E3D-IQA can effectively diagnose and improve the interpretability of MLIP energy allocations using partially labeled datasets.
Intra-atomic supervision is crucial for achieving accurate energy decomposition in machine-learning interatomic potentials, revealing hidden allocation failures.
Machine-learning interatomic potentials (MLIPs) can reproduce potential energies and forces accurately, but their internal energy allocation is often difficult to interpret. E3D-IQA is introduced as a diagnostic framework connecting the latent edge-energy representation of an Allegro-type MLIP with Interacting Quantum Atoms (IQA) energy decomposition. The Allegro edge-energy path is retained as a latent pair contribution, while a node-energy path is trained against the IQA intra-atomic energy. IQA interatomic energies are not direct training targets; instead, the learned edge energies are evaluated after training against the IQA pair terms. Tests on H/C/N/O organic reaction structures show that intra-atomic supervision is essential: energy and force training alone does not recover an IQA-like one-body/two-body allocation. With intra-atomic supervision, node energies reproduce IQA intra-atomic terms, and latent edge energies show meaningful correspondence with IQA interatomic terms. Residual errors are concentrated in positive or weak pair interactions, exposing internal allocation failures that remain hidden in total-energy and force metrics. Adding structures labeled only with energies and forces improves transfer to larger molecules and reduces decomposition errors. E3D-IQA therefore provides a route for diagnosing and guiding latent scientific representations using partially labeled quantum-chemical datasets.