← Latest papers
🔬 materials science

Systematic Evaluation of Training Data Generation Strategies for Machine Learning-Based Molecular and Materials Property Prediction

This paper systematically evaluates various training data acquisition strategies for machine learning-based molecular and materials property prediction, demonstrating that while uncertainty-based sampling generally outperforms other methods, the optimal strategy ultimately depends on the specific interplay between the model, dataset, and prediction task.

Original author: Minh Nghia Nguyen, Beate Paulus, Akhil S. Nair

Published 2026-11-13
📖 4 min read☕ Coffee break read

Original author: Minh Nghia Nguyen, Beate Paulus, Akhil S. Nair

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the quest to discover new medicines, stronger batteries, or more efficient solar cells, scientists often face a daunting bottleneck: the need to know the properties of countless chemical structures before they can be built or tested. Traditionally, figuring out how a molecule behaves or how stable a crystal is requires running complex computer simulations or performing expensive laboratory experiments. These processes are slow and consume vast amounts of computing power. To speed things up, researchers have turned to machine learning, a type of artificial intelligence that can learn to predict these properties by studying examples. However, for these computer models to learn effectively, they need a massive amount of training data. The central challenge is that not all data is equally useful. Just as reading the same book ten times teaches you nothing new, feeding a model thousands of nearly identical chemical structures wastes time and resources. The goal is to find the most informative examples to teach the model, rather than just collecting as many as possible.

A team of researchers set out to solve this problem by systematically testing different ways to choose which chemical examples to include in a training set. They asked a simple but critical question: if you have a limited budget for running expensive calculations, what is the best strategy to pick the data that will make your prediction model the most accurate? They compared several approaches, ranging from picking examples completely at random to using more sophisticated methods that look for uncertainty or diversity in the chemical space. The study focused on two main types of materials: small organic molecules and inorganic crystals, which are the building blocks of many solid materials. They tested these strategies on four different datasets, varying in size from a few thousand to over one hundred thousand examples, and used two different types of machine learning models: one that relies on pre-defined chemical descriptions and another that learns directly from the atomic structure.

The researchers found that the best way to gather data depends heavily on the specific task, the type of material, and the model being used. There is no single "magic bullet" strategy that works for every situation. However, one approach consistently outperformed the others across the board: a method that prioritizes examples where the model is most unsure of its answer. By focusing on these uncertain cases, the model learns to correct its biggest mistakes first, rather than wasting time on examples it already understands well. This "uncertainty sampling" strategy allowed the models to reach high levels of accuracy using significantly fewer data points than if they had simply picked examples at random. In some cases, this approach reduced the number of required examples by nearly a third, a massive saving when each data point represents a costly computer simulation.

The study also revealed that the size of the available data pool matters greatly. When the pool of potential examples is small and diverse, such as in the case of complex crystal stability calculations, the benefit of smart selection becomes even more pronounced. In these data-scarce scenarios, random selection often fails to capture the full range of chemical possibilities, whereas the uncertainty-based method ensures the model sees the most critical variations. Conversely, in very large datasets containing many redundant or similar structures, even random selection can eventually find enough useful information, though the smart method still gets there faster. The researchers also discovered that the choice of the machine learning model itself is just as important as the data selection strategy. A model that understands the detailed 3D arrangement of atoms performed far better than one that relied only on simplified chemical descriptions, regardless of how the data was chosen.

Ultimately, the work provides a clear guide for scientists who need to build predictive models without wasting resources. It shows that while the specific best strategy can vary, relying on the model's own uncertainty to guide data collection is a robust and efficient path forward. This is particularly valuable for properties that are expensive to calculate, where saving even a few hundred calculations can translate into significant time and money saved. The findings suggest that the future of materials discovery lies not just in having more data, but in having the right data, selected with a strategy that respects the unique needs of the model and the material being studied. By balancing the choice of model with a smart data acquisition strategy, researchers can accelerate the discovery of new materials without the need for exhaustive and costly trial and error.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →