TabPFN Predicts Energetic Material Properties Without Tuning

A tabular foundation model called TabPFN predicted three key properties of energetic materials more accurately than the best previously published models, and it did so without any hyperparameter tuning, according to an arXiv preprint by researchers at the Norwegian Defence Research Establishment (FFI) and the Norwegian Computing Center (NR). The finding matters because several of these properties can only be measured through hazardous and expensive experiments, so more reliable predictions could inform how explosives are produced, transported, stored and handled.

What did the study set out to predict?

The team focused on three properties central to assessing energetic materials: impact sensitivity, decomposition temperature and crystal density. Broadly, impact sensitivity describes how readily a material reacts to a mechanical shock, decomposition temperature is the point at which it breaks down, and crystal density reflects how tightly its molecules pack. Each normally demands careful laboratory work to pin down.

To predict them, the researchers compared a slate of models. Alongside TabPFN they tested five established approaches meant to represent the most widely used methods in earlier work: the lasso, random forests, XGBoost, feed-forward neural networks and Gaussian process regression. They also drew on Chemprop, a message-passing neural network that turns a molecule's structure into a learned numerical representation.

How did TabPFN compare with the other models?

According to the preprint, TabPFN came out ahead across the board. It beat the best published models for all three properties, and it substantially outperformed the established machine-learning baselines, including a tuned version of the gradient-boosting method XGBoost, random forests and Gaussian processes. The advantage held regardless of how the molecules were represented: whether the model received hand-crafted chemical descriptors or the learned embeddings produced by Chemprop.

One result stands out. When TabPFN was fed Chemprop's own embedding, it was more accurate than Chemprop's own predictions, for all three properties. In effect, the general-purpose tabular model made better use of the specialised network's representation than the network did itself. Speed was another benefit: TabPFN was among the fastest models tested and ran more than an order of magnitude faster than XGBoost.

What is TabPFN, and why does it suit this kind of data?

TabPFN is a tabular foundation model introduced by Hollmann and colleagues in 2025. Rather than being trained to predict one specific target, it was pre-trained on a large collection of purely synthetic datasets. It then makes predictions on a new problem through in-context learning, using the application's own training examples as context. That design means it needs no additional training or fine-tuning to tackle a fresh dataset, though its weights can be fine-tuned if desired. The latest version, v3.5, can handle datasets with up to one million rows or 20,000 columns.

That profile fits the problem well. Energetic-material datasets tend to be modest in size and built from heterogeneous features — some integer-valued, such as atomic counts, some real-valued, such as oxygen balance, and some categorical, such as the identifier of a molecule's most sensitive functional group. For data like this, tree-based methods such as random forests and XGBoost have historically outperformed deep neural networks, which is part of why the authors saw a general tabular model as a strong candidate.

Does it report uncertainty and explain its predictions?

Yes. Because TabPFN predicts the entire distribution of the target rather than a single point estimate, it offers built-in uncertainty quantification. The researchers compared these predicted distributions against the empirical spread of compounds that had been measured repeatedly, a check on whether the model's confidence is well calibrated.

For interpretability, the team applied Shapley values, a model-agnostic tool from explainable AI that ranks the input features by how much each contributes to a prediction. The appendix works through this analysis for four well-known compounds — HMX, RDX, NQ and TNT — showing which descriptors drive the model's estimates for each.

Why is predicting these properties so difficult?

The authors are candid about the obstacles. Measuring these properties can be dangerous and costly, and many laboratories keep tight restrictions on sharing data, so there are few openly available, high-quality datasets of even moderate size. Chemists also use different definitions of the same target, which makes it hard to pool datasets together. The authors add that neural networks applied to this problem have typically used only a few hidden layers and have not drawn on developments such as in-context learning or transformer architectures — the very ideas behind TabPFN's design. Together with the heterogeneous chemical descriptors, these factors have long made the task awkward for conventional deep learning, and help explain why simpler tree-based models had often held the upper hand.

How can you explore this with Grafy?

If you want to read the study closely, Reader can load the arXiv paper and let you question it directly — asking, for instance, how the Shapley analysis was computed or how the two molecule representations differ. To compare how different AI models summarise the method or debate its limitations, you can put the same question to several of them side by side in Chat. And because TabPFN is a Python library, Grafy Code is a place to write and run code that loads the model and tries a small benchmark of your own. The specific study concerns energetic materials, but the broader lesson — that a foundation model can beat tuned baselines on small, mixed-type tables — carries over to many tabular problems.

Frequently asked questions

Did TabPFN beat tuned XGBoost in the study?

Yes. According to the preprint, TabPFN substantially outperformed a tuned version of XGBoost, along with random forests and Gaussian processes, on all three properties. It also ran more than an order of magnitude faster than XGBoost. The authors note this required no hyperparameter tuning of TabPFN itself, since the model relies on in-context learning rather than being trained on the target dataset.

What three properties did the researchers predict?

The study predicted impact sensitivity, decomposition temperature and crystal density of energetic materials. These properties matter for the safe production, transport, handling, storage and destruction of explosives, yet they often require lengthy, hazardous and costly experiments to measure. The researchers reported that TabPFN outperformed the best published models for all three, using either hand-crafted descriptors or Chemprop's learned embeddings.

What is TabPFN and how does it make predictions?

TabPFN is a tabular foundation model introduced by Hollmann and colleagues in 2025. Instead of being trained on a specific target, it was pre-trained on many purely synthetic datasets and then predicts on new problems through in-context learning, using the provided training rows as context. The latest version, v3.5, handles up to one million rows or 20,000 columns, and the model needs no fine-tuning to use.

---

Source: Grafy Research Lab — Accurate predictions of properties of energetic materials with tabular foundation models