Differential Learning for Robust Prediction of Thermal Stability with Application to Energetic Materials

Every material has a breaking point, a moment when heat becomes too much to bear and the chemical bonds holding it together begin to snap. For the scientists who design energetic materials—substances like explosives and propellants that power rockets or drive machinery—knowing this breaking point is a matter of life and death. If a material is too unstable, it might explode prematurely in storage; if it is too stable, it might fail to perform its job when needed. The standard way to measure this safety margin is to heat a sample in a lab and record the exact temperature at which it starts to decompose. However, this simple measurement is surprisingly difficult to trust. Different laboratories use different heating speeds, different amounts of sample, and different ways of reporting the results. When researchers try to gather these numbers from many different sources to build a computer model, the data becomes a noisy, inconsistent mess, making it nearly impossible to teach a machine to predict which new molecules will be safe and which will be dangerous.

A team of researchers at Los Alamos National Laboratory and Texas Tech University has found a clever way to cut through this noise. Instead of trying to teach a computer to predict the exact temperature at which a molecule will break, they taught it to compare two molecules at a time and decide which one is more stable. This approach, which they call differential learning, ignores the confusing differences between laboratories and focuses entirely on the relative order of the materials. By training artificial intelligence to look at pairs of molecules and predict the difference in their stability, the team created a system that can rank thousands of compounds with high accuracy. Their work suggests that while we cannot always trust a single number from a lab report, we can reliably trust the relationship between two different chemicals.

The researchers applied this method to a dataset of organic molecules containing only carbon, hydrogen, nitrogen, and oxygen, a common group in energetic materials. They gathered experimental data from published studies, but as expected, the numbers were inconsistent. When they tried to use standard computer models to predict the exact decomposition temperature for each molecule, the results were poor. The models struggled to distinguish between molecules that were very similar and often failed to identify the most stable or least stable outliers. This failure highlighted the problem: the noise in the data was drowning out the true chemical signals.

To solve this, the team changed the question. They stopped asking the computer, "What is the temperature of this molecule?" and started asking, "Is this molecule more stable than that one?" They fed the computer pairs of molecules and asked it to predict the difference in their stability. This shift allowed them to generate a massive number of training examples from a relatively small set of data. If they had 850 molecules, they could create hundreds of thousands of unique pairs to study. The computer learned to spot the subtle structural features that made one molecule hold together better than another, effectively filtering out the experimental errors that plagued the absolute numbers.

They tested two different types of computer models to see which worked best. One model looked at the molecules as simple graphs of atoms and bonds, learning directly from the shape of the molecule. The other model used a list of calculated properties, such as how much energy it would take to break a specific bond, to make its decision. Both models performed remarkably well. When asked to rank the molecules in a test set, they correctly identified the more stable member of a pair about 87 percent of the time. This was a significant improvement over the traditional methods, which had struggled to make sense of the same data. The success of both models, despite using different ways of looking at the molecules, gave the researchers confidence that they had found a robust way to handle messy experimental data.

Beyond just ranking the molecules, the researchers wanted to understand what the computer was actually learning. They used a technique to peek inside the "black box" of the neural network and see which features were driving the predictions. They found that the computer relied heavily on two main factors: the balance of oxygen in the molecule and the strength of its weakest chemical bonds. This aligns with what chemists have long suspected. In energetic materials, the presence of oxygen often correlates with higher performance but lower stability, while the weakest bond in a chain acts as a trigger for decomposition. However, the study also revealed that no single factor tells the whole story. The computer did not just look at the weakest bond in isolation; it considered the entire combination of the molecule's shape, its bond strengths, and its chemical makeup. This suggests that predicting stability is a complex puzzle where the interaction of many small parts matters more than any single piece.

To demonstrate how this insight could be used in the real world, the team used a famous, highly stable molecule called TATB as a starting point. They asked the computer to imagine small changes to the molecule's structure and predict how those changes would affect its stability. The model correctly identified that replacing certain groups of atoms with others could either strengthen or weaken the molecule. For instance, it predicted that swapping out specific hydrogen-containing groups for oxygen-containing ones would make the molecule significantly less stable. Conversely, it suggested that removing certain unstable groups entirely could make the molecule even more robust. These findings provide a clear guide for chemists who want to design new materials: they know which structural features to keep and which to avoid to achieve the right balance of power and safety.

The study concludes that this differential approach is a powerful tool for materials discovery, especially when the available data is imperfect. By focusing on relative differences rather than absolute values, the researchers were able to build a model that works well despite the inconsistencies in how the data was collected. They have made their dataset and their models available to the public, hoping that others will use this method to design safer and more effective materials. While the current models are based on the properties of isolated molecules in a vacuum, the researchers acknowledge that real-world materials exist in solid forms where molecules interact with each other. Future work will need to account for these interactions, but the success of this new method offers a promising path forward for navigating the complex chemistry of thermal stability.