Research map / Image generation / Autoregressive and latent generation

AMGenC: Generating Charge Balanced Amorphous Materials

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Equivariant diffusion for 3D molecules · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AMGenC, a flow-matching generative inverse design method for amorphous materials that guarantees charge-balanced element assignments via OT-coupled element noise, per-step soft Gauss-Newton projection, and a final DP-based discrete projection.

Problem

Generative inverse design of amorphous materials often produces charge-unbalanced samples because element assignments are unconstrained and the charge constraint is non-differentiable; post-hoc filtering discards >90% of samples for large cells.

Approach

Combines three components: element noise centered on the training element distribution (OT-coupled), a per-step Gauss-Newton projection using a softmax-relaxed charge gradient with exact discrete charge value, and a final minimum-cost discrete swap via dynamic programming to guarantee Q=0.

Key results

• AMGenC achieves 100% P(Q=0) on all four configurations, versus at most 6.8% for baselines (a-SiO2) and at most 2.0% (MEG).
• On a-SiO2 | G, AMGenC reaches MAE 3.32 and MAPE 14.1%, ranking first on all metrics; on MEG it trades top position with AMDEN.
• Time to 1,000 charge-balanced samples on MEG drops to about 47 minutes versus roughly 39 to 49 hours for the best baseline.

Surprising findings

• The per-step soft projection also improves inverse design metrics (e.g., E MAE 10.68 to 9.95 in the ablation), which the authors attribute to alignment with the all-charge-balanced training distribution.

Limitations

• Generated samples are computational predictions and not directly synthesizable experimentally.
• Generated samples extrapolate more readily to lower property values than higher ones, bounded by physical constraints.
• Radial distribution functions of generated samples are sometimes lower and broader than training data.

Experimental setup

Datasets: a-SiO2 dataset (6,000 MD-generated samples) MEG dataset (9,027 glass samples)

Benchmarks: a-SiO2 | G a-SiO2 | RSD MEG | E+C_Li MEG | G

Metrics: MAE RMSE MAPE P(Q=0) mean absolute charge charge std wall-clock time to 1k charge-balanced samples

Baselines: CDVAE CrystalFlow MatterGen Graphite AMDEN

Builds on: flow matching PCFM EGNN optimal transport coupling

Closest papers by meaning

• LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation (NeurIPS 2025 · similarity 0.87)
• Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.85)
• Co-PiLOT: Constrained Physics-Informed Latent Optimization for Target-Driven Inverse Design (NeurIPS 2026 · similarity 0.85)
• WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry (ICML 2025 · similarity 0.85)
• Generating Symmetric Materials using Latent Flow Matching (NeurIPS 2026 · similarity 0.84)
• SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models (ICLR 2025 · similarity 0.84)