Scientists constantly study phase transitions in various materials and systems to understand their properties better. However, quantifying phase changes in unknown systems with limited data can be challenging. Researchers from MIT and the University of Basel have developed a generative artificial intelligence model to automatically map out phase diagrams for novel physical systems. This physics-informed machine-learning approach is more efficient than manual techniques and does not require large labeled training datasets, making it a valuable tool for researchers investigating the thermodynamic properties of materials or detecting entanglement in quantum systems.
The hope is that this approach can help scientists discover unknown phases of matter autonomously, providing a pipeline for automated scientific discovery of new exotic properties of phases. By using the generative model, researchers can scan large new systems in an automated way and pinpoint important changes in the system. This technique could revolutionize the way scientists study materials and systems, allowing for faster and more accurate identification of phase transitions.
The researchers used the Julia Programming Language, a popular scientific computing language known for its tools to construct generative models. Generative models estimate the probability distribution of data and generate new data points fitting the distribution. By using simulations of physical systems, researchers can model the probability distribution, providing valuable insight into the measurement statistics of the system. This distribution serves as the basis for constructing a generative classifier that can determine the phase the system is in based on parameters like temperature or pressure.
This approach allows for more knowledgeable modeling, going beyond feature engineering on data samples to directly incorporate system knowledge into the classifier. By approximating the probability distributions underlying measurements from the physical system, the classifier can perform better than other machine-learning techniques. It can work automatically without extensive training, significantly enhancing the computational efficiency of identifying phase transitions. Researchers can use this approach for various binary classification tasks in physical systems, such as detecting entanglement in quantum systems or determining the best theory to solve a specific problem.
In the future, researchers plan to study theoretical guarantees regarding the number of measurements needed to detect phase transitions effectively and estimate the computation required. The project was funded by the Swiss National Science Foundation, the MIT-Switzerland Lockheed Martin Seed Fund, and MIT International Science and Technology Initiatives. By leveraging artificial intelligence and generative models, researchers hope to advance the field of studying phase transitions in novel materials and systems, unlocking new possibilities for scientific discovery and exploration of unknown phases of matter.