Machine learning simulations strengthen the liquid-liquid transition hypothesis in water

The properties of liquid water cooled below its freezing temperature (supercooled water) and of cold, non-crystalline solid forms of water (glassy water, also referred to as amorphous ice) have long attracted the attention of physicists, chemists, and materials scientists. Computer simulations have played a central role in advancing our understanding of these states of water. In particular, molecular simulations allow researchers to avoid crystallization, making it possible to study metastable liquid water without the interference of freezing. In 1992, molecular dynamics simulations carried out by Francesco Sciortino, one of the review co-authors, and collaborators led to the formulation of the liquid–liquid phase transition hypothesis. According to this scenario, the unusual properties of water, such as the sharp increase in the isothermal compressibility and isobaric heat capacity upon cooling, reflect an underlying metastable first-order phase transition between two distinct liquid phases in deeply supercooled water, which terminates at a liquid–liquid critical point.

The hypothesis became the subject of intense debate in 2011, when the very existence of the liquid–liquid transition was challenged by conflicting results from simulations carried out by researchers at Princeton University and the University of California, Berkeley. The disagreement was definitively resolved in 2018, when Pablo Debenedetti, one of the review’s co-authors, and collaborators showed that it originated from a problem with one of the simulation codes. Since then, research has focused on characterizing the transition, precisely locating the critical point, and exploring its consequences for liquid and glassy water.

According to lead author Pablo Piaggi, Ikerbasque Research Fellow in the Theory Group at nanoGUNE, recent advances in machine learning have been particularly important: “Machine-learning-based molecular dynamics simulations have provided new evidence for the liquid–liquid transition obtained entirely from first principles. They have also enabled us to locate the liquid–liquid critical point with unprecedented accuracy.” These new methods combine the accuracy of quantum-mechanical calculations with the computational efficiency needed to investigate the complex behaviour of deeply supercooled water. 

The review article describes the latest advances in the computational study of supercooled and glassy water, highlighting how increasingly sophisticated approaches, particularly advanced atomistic simulation methods and machine learning, are opening new avenues for understanding cold, disordered states of water that are metastable with respect to crystalline phases.