Out-of-Distribution Inverse Design of Elastic Networks with Differentiable Graph Neural Network Molecular Dynamics
Abstract
Machine-learning-based inverse design can accelerate the discovery of materials with targeted properties, but conventional structure--property models often require large training datasets and generalize poorly beyond their training distribution. Here, we present a differentiable inverse design framework based on a graph neural network molecular dynamics simulator. By combining a short dynamical initialization with physics-based refinement during simulation, the framework enables optimization well beyond the conditions represented in the training data. Using disordered elastic networks, we show that a simulator trained only on non-auxetic systems with Poisson's ratios between 0.1 and 0.4 can design strongly auxetic networks with values as low as -0.3. The framework also produces stress--strain responses outside the range observed during training, creates localized mechanical defects absent from the training data, and generalizes across system size, enabling optimization of networks tested up to 5000 nodes despite being trained only on networks with fewer than 200 nodes.
Related
- Differentiable library-based inverse design of achromatic metalens for full-color near-eye displays
- Dopant-Controlled Gas Adsorption and Electronic Response in AlN-and BP-Doped Two-Dimensional Fullerene Networks: A DFT and Ab Initio Molecular Dynamics Study
- Differentiable mass transfer modeling enables wet lab validated inverse analysis of liquid chromatography
- MatUQ: a benchmark for uncertainty-aware out-of-distribution materials property prediction with graph neural networks
- Thermodynamic and Molecular Characterization of Adsorption on Zeolites: A Unified Framework Combining Inverse Gas Chromatography, Hamaker Theory, and Nonlinear Lewis Acid–Base Modeling
BibTeX
@misc{shteingolts2026outofdistribution,
title = {{Out-of-Distribution Inverse Design of Elastic Networks with Differentiable Graph Neural Network Molecular Dynamics}},
author = {Sergey A. Shteingolts and Salman N. Salman and Ron Levie and Dan Mendels},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.06655},
eprint = {2609.06655},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.06655},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - UNPB
TI - Out-of-Distribution Inverse Design of Elastic Networks with Differentiable Graph Neural Network Molecular Dynamics
AU - Sergey A. Shteingolts
AU - Salman N. Salman
AU - Ron Levie
AU - Dan Mendels
PY - 2026
DA - 2026/09/06
PB - arXiv
DO - 10.48550/arxiv.2609.06655
UR - https://arxiv.org/abs/2609.06655
AB - Machine-learning-based inverse design can accelerate the discovery of materials with targeted properties, but conventional structure--property models often require large training datasets and generalize poorly beyond their training distribution. Here, we present a differentiable inverse design framework based on a graph neural network molecular dynamics simulator. By combining a short dynamical initialization with physics-based refinement during simulation, the framework enables optimization well beyond the conditions represented in the training data. Using disordered elastic networks, we show that a simulator trained only on non-auxetic systems with Poisson's ratios between 0.1 and 0.4 can design strongly auxetic networks with values as low as -0.3. The framework also produces stress--strain responses outside the range observed during training, creates localized mechanical defects absent from the training data, and generalizes across system size, enabling optimization of networks tested up to 5000 nodes despite being trained only on networks with fewer than 200 nodes.
ER -