[Paper Review] Adversarial reverse mapping of condensed-phase molecular structures: Chemical transferability

Marc Stieffenhofer, Tristan Bereau+1 more|arXiv (Cornell University)|Jan 13, 2021

Machine Learning in Materials Science30 references28 citations

TL;DR

This paper introduces deepBackmap (DBM), a conditional generative adversarial network that enables chemical transferability in reverse mapping of condensed-phase molecular structures. By training solely on small-molecule liquids (octane and cumene), DBM successfully reconstructs high-fidelity atomistic structures of complex polymers like syndiotactic polystyrene (sPS), achieving near-identical pair correlation functions and accurate Lennard-Jones energy distributions, demonstrating robust generalization despite limited chemical overlap.

ABSTRACT

Switching between different levels of resolution is essential for multiscale modeling, but restoring details at higher resolution remains challenging. In our previous study we have introduced deepBackmap: a deep neural-network-based approach to reverse-map equilibrated molecular structures for condensed-phase systems. Our method combines data-driven and physics-based aspects, leading to high-quality reconstructed structures. In this work, we expand the scope of our model and examine its chemical transferability. To this end, we train deepBackmap solely on homogeneous molecular liquids of small molecules, and apply it to a more challenging polymer melt. We augment the generator's objective with different force-field-based terms as prior to regularize the results. The best performing physical prior depends on whether we train for a specific chemistry, or transfer our model. Our local environment representation combined with the sequential reconstruction of fine-grained structures help reach transferability of the learned correlations.

Motivation & Objective

• To investigate chemical transferability in machine learning-based backmapping of molecular systems.
• To assess whether local correlations learned from simple liquids can generalize to complex polymers like syndiotactic polystyrene (sPS).
• To evaluate the impact of physics-based priors on model performance and transferability.
• To develop a scalable, autoregressive deep learning framework for high-resolution reconstruction of condensed-phase systems.

Proposed method

• Trained a conditional generative adversarial network (cGAN) on coarse-grained (CG) configurations of octane and cumene liquids.
• Used a voxel-based 3D representation with multi-channel features encoding molecular topology and spatial relationships.
• Implemented an autoregressive generation strategy, reconstructing atoms sequentially based on local environment context.
• Incorporated physics-based priors into the generator loss: energy-minimizing prior (p1) and energy-matching prior (p2).
• Employed a local environment representation to capture short-range interactions critical for transferability.
• Used k-means clustering and confusion matrices on low-dimensional projections to assess conformational space coverage and structural fidelity.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on small-molecule liquids generalize to backmap complex polymer melts without retraining?
• RQ2 How do different physics-based priors affect the quality and transferability of generated atomistic structures?
• RQ3 To what extent do local atomic environments in small molecules transfer to larger, chemically distinct systems?
• RQ4 Does the autoregressive, local-environment-based architecture preserve statistical and energetic properties in transferred systems?
• RQ5 How well does the model reproduce key structural and energetic distributions (e.g., pair correlation, Lennard-Jones energy) in the target system?

Key findings

• The model trained on octane and cumene successfully backmapped sPS with near-identical pair correlation functions to the reference, indicating high structural fidelity.
• Lennard-Jones energy distributions in the backmapped sPS structures closely matched the reference, with minimal deviation.
• The confusion matrix analysis showed that while cluster assignments did not perfectly align, the relative populations of conformational clusters were well reproduced, indicating correct statistical sampling.
• The energy-minimizing prior (p1) yielded the highest-quality structures in the chemically transferred setting, though it led to unrealistically low energies when trained on sPS.
• The energy-matching prior (p2) and no prior performed comparably, suggesting that data-driven learning without explicit physics priors can be sufficient for transferability.
• The model struggled with long-range or orientation-specific features, such as phenyl ring orientation relative to the backbone, indicating limits in generalizing complex steric and electronic effects.