Machine Learning Algorithms for Predicting the Thermodynamic Stability of Molecular Crystals
Project ID: 2531bd1695
(You will need this ID for your application)
Research Theme: Artificial Intelligence and Robotics
Research Area(s):
physics
materials
engineering
UCL Lead department: Physics and Astronomy
Lead Supervisor: Venkat Kapil
Project Summary:
Why this research is important Molecular crystals play a central role in pharmaceuticals, fertilisers, food, and advanced materials. Their properties depend sensitively on which crystalline form — or polymorph — is most stable under given conditions. Predicting this stability is one of the most challenging unsolved problems in materials modelling: the free energy differences between polymorphs is below the accuracy of standard computational methods. A reliable, quantum-level framework for polymorph stability would transform our ability to design molecular materials. This “hard science” problem is of growing interest to emerging AI startups, the pharmaceutical industry, and big tech companies.
Who you will be working with You will join a collaboration between the Kapil group (UCL Physics) and the Salvalaglio group (UCL Chemical Engineering). The Kapil group specialises in machine learning interatomic potentials and quantum thermodynamics, while the Salvalaglio group develops advanced free energy and sampling methods. Together, they provide a unique interdisciplinary environment at the physics–engineering interface.
What you will be doing You will develop and apply two complementary machine learning approaches to overcome the accuracy–cost barrier in crystal structure prediction. First, you will fine-tune foundation machine-learned interatomic potentials on small, system-specific datasets to achieve quantum-accurate forces efficiently. Second, you will implement generative reweighting methods, based on normalizing flows, to estimate free energy differences between polymorphs from limited sampling. Combined, these strategies will deliver the first practical workflow for thermodynamically rigorous polymorph ranking under realistic conditions.
Who we are looking for We seek an enthusiastic student with a strong background in physics, chemistry, materials science, or a related discipline. Prior experience in molecular simulation, statistical mechanics, or machine learning is an advantage. Most importantly, you should be motivated to develop new computational methods and excited by the challenge of applying machine learning to physical science problems.