Conditional latent distillation enables controllable molecular generation in data-limited soft materials
- 1Department of Chemistry
- 2Louisiana State University
- 3Department of Chemical Engineering
Abstract
De novo molecular generation is particularly difficult when materials-specific datasets are too small to support reliable training of a new generative backbone. We develop a framework for controllable molecular generation based on conditional latent distillation and chemistryconstrained construction. A variational autoencoder (VAE) pretrained on more than 20,000 polymer molecules supplies a starting molecular representation, reducing reliance on training a new generator from approximately 1,600 surfactant or 200 anion-exchange-membrane (AEM) records. A compact conditional student learns to predict teacher latent coordinates from molecular descriptors and property-associated inputs, while molecular exploration and chemistry-aware editing govern which candidate structures are constructed and retained. In the primary surfactant application, Bayesian optimization produced 841 cleaned novel structures and retained 118, compared with 205 and 15 from nearest-neighbor exploration under equal 20,000-attempt mutation budgets. An independent distillation-guided teacher-latent retrieval benchmark yielded 2.32 more retained candidates per request than anchor-count-matched random selection (95% confidence interval (CI), 1.78–2.86). The conditional mapping also reduced teacher-coordinate prediction error by 40.3% and decoderdistribution divergence by 59.1% relative to the strongest tested regression baselines, validating the distillation-based interface. Molecular dynamics characterized selected surfactants at a water–decane interface. We then applied the chemistry-constrained construction and screening workflow to AEM-related building blocks, obtaining 125 screened candidates and assessing selected cations in water. These results demonstrate controllable molecular exploration and structural selection in data-limited settings, without claiming experimentally verified control of final materials properties.
Related
- Graph Symbolic Distillation: Decoding Graph Data into Interpretable Formulas across Materials and Chemistry
- MolMiner: Toward Controllable, Three‐Dimensional‐Aware, Fragment‐Based Molecular Design
- Bio-Based Smart Packaging Materials for Next-Generation Food Systems
- Review: next-generation metal hydride systems for sustainable hydrogen storage—kinetics, nanostructuring, and intelligent materials discovery
- Next-generation electronic packaging materials: thermal-dielectric properties and design strategies in PI and PTFE-based composites
BibTeX
@misc{das2026conditional,
title = {{Conditional latent distillation enables controllable molecular generation in data-limited soft materials}},
author = {Debojyoti Das and Pegah Naghshnejad and Toheeb O. Balogun and Gabriela C. Theis Marchan and Revati Kumar and José A. Romagnoli},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010751/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010751/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Conditional latent distillation enables controllable molecular generation in data-limited soft materials
AU - Debojyoti Das
AU - Pegah Naghshnejad
AU - Toheeb O. Balogun
AU - Gabriela C. Theis Marchan
AU - Revati Kumar
AU - José A. Romagnoli
PY - 2026
DA - 2026/10/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010751/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010751/v1
AB - De novo molecular generation is particularly difficult when materials-specific datasets are too small to support reliable training of a new generative backbone. We develop a framework for controllable molecular generation based on conditional latent distillation and chemistryconstrained construction. A variational autoencoder (VAE) pretrained on more than 20,000 polymer molecules supplies a starting molecular representation, reducing reliance on training a new generator from approximately 1,600 surfactant or 200 anion-exchange-membrane (AEM) records. A compact conditional student learns to predict teacher latent coordinates from molecular descriptors and property-associated inputs, while molecular exploration and chemistry-aware editing govern which candidate structures are constructed and retained. In the primary surfactant application, Bayesian optimization produced 841 cleaned novel structures and retained 118, compared with 205 and 15 from nearest-neighbor exploration under equal 20,000-attempt mutation budgets. An independent distillation-guided teacher-latent retrieval benchmark yielded 2.32 more retained candidates per request than anchor-count-matched random selection (95% confidence interval (CI), 1.78–2.86). The conditional mapping also reduced teacher-coordinate prediction error by 40.3% and decoderdistribution divergence by 59.1% relative to the strongest tested regression baselines, validating the distillation-based interface. Molecular dynamics characterized selected surfactants at a water–decane interface. We then applied the chemistry-constrained construction and screening workflow to AEM-related building blocks, obtaining 125 screened candidates and assessing selected cations in water. These results demonstrate controllable molecular exploration and structural selection in data-limited settings, without claiming experimentally verified control of final materials properties.
ER -