ToSCA: Total Scattering Conditioned Variational Autoencoder for Atomic Structure Prediction from Pair Distribution Function Data
- 1Department of Chemistry & Nano-Science Center
- 2University of Copenhagen
- 3Department of Computer Science
Abstract
Designing new materials is often slow and experimentally intensive, with structural characterization representing a particular bottleneck. Recent advances in machine learning (ML) aim to accelerate this process by predicting plausible atomic structures describing a material directly from experimental data. Here we present the total scattering conditioned variational autoencoder (ToSCA), a generative ML model for predicting plausible atomic structures of metal oxide nanomaterials from total scattering and pair distribution function (PDF) data. ToSCA learns a chemically meaningful latent representation and predicts unit cells within tolerances suitable for subsequent real-space refinement. The model also shows behavior consistent with learning structural relationships between related crystal types, and can correctly place the new CHILI-Interpolation dataset in the latent space. Finally, we demonstrate that ToSCA is able to retain useful performance in a realistic use-case by testing it on 11 experimental PDFs from 3 crystal types, including test cases outside of the training distribution.
Related
- Inverse design of multi-band terahertz metamaterial sensors with nested topologies via a transformer-conditioned variational autoencoder
- Gas-Conditioned Message Passing Graph Neural Network for Permeability Prediction in Polymeric Membranes
- Logarithmic-scale variational quantum eigensolver for off-lattice protein structure prediction in continuous torsional angle space
- Evaluating a Lightweight Neural Network for Aggregate Isotope Distribution Prediction
- When Do Quantum-Chemical Descriptors Improve Aqueous Solubility Prediction? A Residual-Rescue Benchmark under Chemical Distribution Shift
BibTeX
@misc{friisjensen2026tosca,
title = {{ToSCA: Total Scattering Conditioned Variational Autoencoder for Atomic Structure Prediction from Pair Distribution Function Data}},
author = {Ulrik Friis-Jensen and Frederik L. Johansen and Erik B. Dam and Raghavendra Selvan and Kirsten M. Ø. Jensen},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15009767/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009767/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - ToSCA: Total Scattering Conditioned Variational Autoencoder for Atomic Structure Prediction from Pair Distribution Function Data
AU - Ulrik Friis-Jensen
AU - Frederik L. Johansen
AU - Erik B. Dam
AU - Raghavendra Selvan
AU - Kirsten M. Ø. Jensen
PY - 2026
DA - 2026/10/02
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009767/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009767/v1
AB - Designing new materials is often slow and experimentally intensive, with structural characterization representing a particular bottleneck. Recent advances in machine learning (ML) aim to accelerate this process by predicting plausible atomic structures describing a material directly from experimental data. Here we present the total scattering conditioned variational autoencoder (ToSCA), a generative ML model for predicting plausible atomic structures of metal oxide nanomaterials from total scattering and pair distribution function (PDF) data. ToSCA learns a chemically meaningful latent representation and predicts unit cells within tolerances suitable for subsequent real-space refinement. The model also shows behavior consistent with learning structural relationships between related crystal types, and can correctly place the new CHILI-Interpolation dataset in the latent space. Finally, we demonstrate that ToSCA is able to retain useful performance in a realistic use-case by testing it on 11 experimental PDFs from 3 crystal types, including test cases outside of the training distribution.
ER -