Brian (Po-Yen) Tung

Inferring optimal solutions from limited data is considered the ultimate goal in scientific discovery. Artificial intelligence offers a promising avenue to greatly accelerate this process. Existing methods often depend on large datasets, strong assumptions about objective functions, and classic machine learning techniques, restricting their effectiveness to low-dimensional or data-rich problems. We present a deep active learning pipeline that combines deep neural networks with a novel tree search to find superior solutions in high-dimensional complex problems with non-cumulative objectives and limited data availability. Our pipeline iteratively approaches the optimum using a neural surrogate and introduces new search mechanisms to bypass the local optimum and minimize the number of samples needed to achieve superior solutions. These contributions enable our pipeline to achieve superior solutions across diverse problems with up to 2,000 dimensions, whereas existing methods are limited to 100 dimensions and require up to 10 times more data points. Our pipeline demonstrates wide applicability, discovering superior solutions in various domain science problems. This advancement enables data-efficient knowledge discovery and opens the path towards scalable self-driving laboratories. Although we focus on problems within the realm of scientific domain, the advancements achieved herein are applicable to a broader spectrum of challenges across all quantitative disciplines.

High-entropy alloys are solid solutions of multiple principal elements that are capable of reaching composition and property regimes inaccessible for dilute materials. Discovering those with valuable properties, however, too often relies on serendipity, because thermodynamic alloy design rules alone often fail in high-dimensional composition spaces. We propose an active learning strategy to accelerate the design of high-entropy Invar alloys in a practically infinite compositional space based on very sparse data. Our approach works as a closed-loop, integrating machine learning with density-functional theory, thermodynamic calculations, and experiments. After processing and characterizing 17 new alloys out of millions of possible compositions, we identified two high-entropy Invar alloys with extremely low thermal expansion coefficients around 2 × 10−6 per degree kelvin at 300 kelvin. We believe this to be a suitable pathway for the fast and automated discovery of high-entropy alloys with optimal thermal, magnetic, and electrical properties.

The application of three-dimensional (3D) imaging techniques, such as X-ray tomography and focussed ion beam scanning electron microscopy (FIB-SEM), is increasingly widespread in microstructural analysis of natural materials. However, our ability to collect high-resolution tomographic datasets, each comprising thousands of two-dimensional (2D) images with millions of pixels, far outstrips our ability to analyse them. Pixel-level segmentation of each 2D image is the first step in any analysis pipeline, but creates a considerable human bottleneck in the workflow that can now be overcome using machine learning. Although advanced pre-trained models such as the Segment Anything Model (SAM) have emerged, conventional segmentation workflows for 3D tomographic data remain limited in comparison. To tackle this, we propose a machine learning workflow that combines SAM with a few-shot learning framework, automating segmentation and minimising user bias. Using SAM, we generate precise annotations from a limited subset of 2D images through basic input prompts, such as points and boxes. These annotations serve as the training data for the few-shot learning model. We benchmark this workflow using a complex 3D FIB-SEM tomographic dataset of the C2 ungrouped carbonaceous chondrite WIS91600. With only 0.6 % of the training data, our method achieves an intersection over union (IoU) score of 80.62 % compared to the ground truth, significantly outperforming widely used methods that achieve a maximum IoU score of 67.07 %. The strong performance on the challenging meteorite dataset highlights its potential for broader application across materials and imaging modalities.

Identification of unknown micro- and nano-sized mineral phases is commonly achieved by analyzing chemical maps generated from hyperspectral imaging data sets, particularly scanning electron microscope—energy dispersive X-ray spectroscopy (SEM-EDS). However, the accuracy and reliability of mineral identification are often limited by subjective human interpretation, non-ideal sample preparation, and the presence of mixed chemical signals generated within the electron-beam interaction volume. Machine learning has emerged as a powerful tool to overcome these problems. Here, we propose a machine-learning approach to identify unknown phases and unmix their overlapped chemical signals. This approach leverages the guidance of Gaussian mixture modeling clustering fitted on an informative latent space of pixel-wise elemental data points modeled using a neural network autoencoder, and unmixes the overlapped chemical signals of phases using non-negative matrix factorization. We evaluate the reliability and the accuracy of the new approach using two SEM-EDS data sets: a synthetic mixture sample and a real-world particulate matter sample. In the former, the proposed approach successfully identifies all major phases and extracts background-subtracted single-phase chemical signals. The unmixed chemical spectra show an average similarity of 83.0% with the ground truth spectra. In the second case, the approach demonstrates the ability to identify potentially magnetic Fe-bearing particles and their background-subtracted chemical signals. We demonstrate a flexible and adaptable approach that brings a significant improvement to mineralogical and chemical analysis in a fully automated manner. The proposed analysis process has been built into a user-friendly Python code with a graphical user interface for ease of use by general users.