Paper review · November 13, 2026
Accurately Predicting Properties of Energetic Materials with a Tabular Foundation Model
Accurate predictions of properties of energetic materials with tabular foundation models
Used without hyperparameter tuning, the tabular foundation model TabPFN outperformed the previous best models at predicting impact sensitivity, decomposition temperature and crystal density of energetic materials.
Problem
Measuring the properties of energetic materials such as explosives requires dangerous and expensive experiments. Public data is scarce, property values come in mixed types (integer, real-valued, categorical), and measurement definitions differ from study to study. Under these conditions, deep learning has often been weaker than tree-based models.
Method
The study uses TabPFN (v3, v3.5), a transformer-based tabular model pretrained on synthetic data. It takes the training data as context and makes predictions in a single forward pass through in-context learning, without updating any weights. Two kinds of input are used: hand-picked molecular descriptors, and learned embeddings from Chemprop's message-passing neural network. It is compared against Lasso, random forest, tuned XGBoost, Gaussian process, a feedforward neural network, and the previous best models for each dataset (VSURF, q-RASPR). Uncertainty is quantified from the predictive distribution, and feature importance is interpreted with Shapley values.
Results
- The authors report that, even without hyperparameter tuning, it was more accurate than the best previously published models (VSURF for impact sensitivity, q-RASPR for the others) on all three properties: impact sensitivity, decomposition temperature and crystal density.
- It outperformed tuned XGBoost, random forest and Gaussian process by a wide margin, and this held for both hand-crafted descriptor and Chemprop embedding inputs.
- On all three properties, TabPFN on top of Chemprop embeddings was more accurate than Chemprop's own predictions.
- It was among the fastest of the compared models and more than 10 times faster than XGBoost.
- The uncertainty from the predictive distribution was compared with the actual distribution of compounds measured repeatedly, and feature importance rankings were given using Shapley values.
Limitations
The datasets are small. Impact sensitivity covers 1,533 measurements of 516 compounds, about a third of which are a single substance, PETN. Decomposition temperature covers 565 compounds. The impact sensitivity data mixes in BAM-method measurements, which use a different definition. Only three properties are covered, and generalization to other properties remains the authors' expectation. The input text was truncated, so specific error figures could not be confirmed.
Why it matters
This is a case showing that a tabular foundation model usable out of the box without tuning can be more accurate and faster than conventional ML, even in a field with little data and dangerous experiments. It could change the default baseline in research settings with many small tabular datasets, such as materials and chemistry.
Key terms
- TabPFN
- A transformer model pretrained on randomly generated synthetic tabular data. It takes a new dataset's training data as context and makes predictions without separate training.
- In-context learning
- A way of predicting a new input by looking only at examples placed in the input context, without changing the model's weights.
- Shapley values
- A method from game theory that fairly divides up how much each input feature contributed to a prediction.
- Chemprop
- An open-source tool that treats molecular structures as graphs and uses a message-passing neural network for embeddings and property prediction.
Original
Authors: Vagn Krister Nes Glomsrud, Mathias Bynke, Lars Henry Berge Olsen, Dennis Christensen · Posted 2026-11-10 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.