Accurate predictions of properties of energetic materials with tabular foundation models
Organizations: Norwegian Defence Research Establishment (FFI) · Norwegian Computing Center (NR)
Abstract
We use TabPFN, a recently developed tabular foundation model, to predict three properties of energetic materials: impact sensitivity, decomposition temperature and crystal density. Without any hyperparameter tuning, TabPFN outperforms the best published models for all three properties. It also substantially outperforms a wide range of established machine-learning models, including tuned gradient boosting (XGBoost), random forests and Gaussian processes. This holds both when the models are given hand-crafted molecular descriptors and when they are given learned representations from Chemprop's message-passing neural network. In particular, for all three properties, TabPFN applied to Chemprop's own embedding is more accurate than Chemprop's predictions. TabPFN is also among the fastest models we consider, more than an order of magnitude faster than XGBoost. As TabPFN predicts the entire distribution of the target rather than just a single value, it offers built-in uncertainty quantification, which we compare with the empirical distributions of compounds with repeated measurements. Finally, we interpret the predictions using Shapley values, which rank the features by their importance.
Figures & tables
| Model | Runtime (s) | ||||
|---|---|---|---|---|---|
| Lasso | 0.324 | (0.004) | 0.580 | (0.010) | 00 1.4 |
| Random forest | 0.307 | (0.004) | 0.624 | (0.011) | 00 1.4 |
| VSURF | 0.304 | (0.004) | 0.631 | (0.010) | 227.7 |
| XGBoost | 0.320 | (0.011) | 0.588 | (0.029) | 0 27.4 |
| Gaussian process | 0.338 | (0.006) | 0.543 | (0.015) | 00 5.1 |
| TabPFN v3 | 0.279 | (0.003) | 0.688 | (0.008) | 00 2.3 |
| Model | Runtime (s) | ||||
|---|---|---|---|---|---|
| Lasso | 0.331 | (0.004) | 0.562 | (0.009) | 0 4.6 |
| Random forest | 0.314 | (0.005) | 0.604 | (0.011) | 0 2.0 |
| XGBoost | 0.320 | (0.024) | 0.588 | (0.065) | 64.6 |
| Gaussian process | 0.369 | (0.004) | 0.456 | (0.012) | 0 5.2 |
| TabPFN v3 | 0.293 | (0.006) | 0.656 | (0.014) | 0 3.1 |
| TabPFN v3.5 | 0.289 | (0.006) | 0.665 | (0.014) | 0 3.4 |
| Model | Runtime (s) | ||||
|---|---|---|---|---|---|
| Lasso | 0.364 | (0.016) | 0.467 | (0.047) | 0 63.2 |
| Random forest | 0.317 | (0.010) | 0.598 | (0.025) | 0 57.2 |
| XGBoost | 0.318 | (0.008) | 0.596 | (0.020) | 271.4 |
| FFNN | 0.337 | (0.015) | 0.544 | (0.042) | 0 54.1 |
| Gaussian process | 0.350 | (0.012) | 0.510 | (0.032) | 0 59.9 |
| TabPFN v3 | 0.306 | (0.010) | 0.626 | (0.025) | 0 57.4 |
| Model | RMSE | MAE | ||||
|---|---|---|---|---|---|---|
| q-RASPR | 50.68 | – | 41.38 | – | 0.676 | – |
| TabPFN v3 | 49.43 | (0.40) | 37.94 | (0.35) | 0.690 | (0.005) |
| TabPFN v3.5 | 49.94 | (0.47) | 39.06 | (0.38) | 0.683 | (0.006) |
| Model | Runtime (s) | ||||
|---|---|---|---|---|---|
| Lasso | 0.0600 | (0.0003) | 0.9025 | (0.0010) | 100.1 |
| Random forest | 0.0422 | (0.0003) | 0.9519 | (0.0006) | 0 73.4 |
| XGBoost | 0.0361 | (0.0004) | 0.9647 | (0.0008) | 405.1 |
| Gaussian process | 0.0916 | (0.0006) | 0.7732 | (0.0029) | 0 75.6 |
| TabPFN v3 | 0.0325 | (0.0002) | 0.9715 | (0.0003) | 0 23.3 |
| TabPFN v3.5 | 0.0314 | (0.0002) | 0.9733 | (0.0003) | 0 31.8 |
| Model | Runtime (s) | ||||
|---|---|---|---|---|---|
| Lasso | 0.0781 | (0.0014) | 0.8351 | (0.0060) | 110.8 |
| Random forest | 0.0358 | (0.0004) | 0.9653 | (0.0008) | 156.1 |
| XGBoost | 0.0351 | (0.0004) | 0.9667 | (0.0007) | 758.9 |
| FFNN | 0.0339 | (0.0005) | 0.9689 | (0.0008) | 0 91.4 |
| Gaussian process | 0.0590 | (0.0011) | 0.9059 | (0.0034) | 176.0 |
| TabPFN v3 | 0.0325 | (0.0003) | 0.9715 | (0.0005) | 117.7 |
| Model | RMSE | MAE | ||||
|---|---|---|---|---|---|---|
| q-RASPR | 0.047 | – | 0.035 | – | 0.939 | – |
| TabPFN v3 | 0.03125 | (0.00007) | 0.02392 | (0.00005) | 0.9737 | (0.0001) |
| TabPFN v3.5 | 0.03041 | (0.00007) | 0.02347 | (0.00004) | 0.9751 | (0.0001) |
| Standalone Chemprop training | Trained Chemprop | Trained TabPFN | Runtime (s) | Better than no training | ||
|---|---|---|---|---|---|---|
| 0.301 (0.009) | 0.636 (0.022) | 23.0 | – | |||
| 0.302 (0.010) | 0.635 (0.024) | 37.3 | 1 5/10 | |||
| 0.301 (0.012) | 0.637 (0.029) | 28.7 | 1 4/10 | |||
| 0.301 (0.010) | 0.637 (0.024) | 29.8 | 1 4/10 | |||
| 0.296 (0.007) | 0.649 (0.018) | 0 1.4 | – | |||
| 0.295 (0.007) | 0.652 (0.018) | 13.8 | 1 9/10 |
| Dataset | TabPFN version | Chemprop encoder | Runtime (s) | Untrained better | ||
|---|---|---|---|---|---|---|
| Impact sensitivity | 3 | Trained | 0.306 (0.010) | 0.626 (0.025) | 57.4 | |
| 3 | Untrained | 0.306 (0.010) | 0.625 (0.025) | 0 3.0 | 1 5/10 | |
| 3.5 | Trained | 0.303 (0.009) | 0.632 (0.023) | 57.2 | ||
| 3.5 | Untrained | 0.295 (0.008) | 0.651 (0.020) | 0 2.9 | 1 8/10 | |
| Decomposition temperature | 3 | Trained | 55.4 (0.8) | 0.614 (0.011) | 19.2 | |
| 3 | Untrained | 53.6 (1.1) | 0.639 (0.015) | 0 1.1 | 10/10 |
| PETN | HMX | RDX | NQ | TNT | |
|---|---|---|---|---|---|
| Observations | 490 | 164 | 130 | 44 | 40 |
| Unique combinations | 8 | 7 | 9 | 3 | 6 |
| Context observations | 1043 | 1369 | 1403 | 1489 | 1493 |
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Hyperparameter | Default | Distribution |
|---|---|---|
| Number of trees | 100 | |
| Learning rate | 0.3 | |
| Maximum tree depth | 6 | |
| Minimum child weight | 1 | |
| Row subsampling rate | 1 | |
| Column subsampling rate per tree | 1 |
Materials Property Prediction
Materials property prediction uses data and physical models to estimate how composition, structure, and processing determine a material’s behavior.
See all 195 papers