Research   

I am currently working on the development of large language models (LLMs) to accelerate MOF synthesis. Here, LLMs have been used for synthesis optimization and to propose new materials based on state-of-the-art methods, thanks to the collaboration with Prof. Joseph Gonzalez . This enabled us to synthesize new MOFs that exhibit some of the best performance for water harvesting. 

One of my main research projects involves training a diffusion probabilistic model for material discovery. Combined with quantum mechanical calculations, machine learning potentials, and force field computations, we were able to generate high-fidelity crystal structures from larger and more complex chemical systems.

Currently, I am involved in various project involving the development of local equivariant neural networks in order to accurately predict quantum mechanical properties. With one of the aim to develop a quantum-inspired Force Fields with high accuracy that could bridge the gap between quantum chemistry accuracy and force field speed up. 

One of my research efforts have been to automate the assignment of Force Fields parameters using databases and on-the-fly ab-initio data. We are developing a Graph-Based Force Fields model to directly derive parameters for the Generalized Amber Force Field (GAFF) from chemical environments and research into the influence of functional forms. This end-to-end parameterization approach eliminates the need for expert-defined procedures and enhances the accuracy and transferability of GAFF across a broader range of molecular complexes. Thus improving to generalize Force Fields to a broader class of systems. 

I have contributed to the development of several high-performance computing software, such as Amsterdam Density Functional (ADF-SCM company), Quantum ESPRESSO, and especially Tinker-HP GPU. Tinker-HP is a massively parallel package (multi-CPUs and multi-GPUs) dedicated to classical molecular dynamics (MD) and to multiscale simulations, using advanced Force Fields ranging from polarizable to classical, as well as neural networks potentials and a framework for handling nuclear quantum effects.

Tinker-HP is able to simulate highly accurate MD simulations on large systems up to millions of atoms at a very high speed. Through collaborative efforts with Carnegie Mellon University and the University of Texas at Austin, we achieved simulations involving tens of millions of atoms using state-of-the-art machine learning potentials on tens of GPUs opening the doors for simulating even larger systems with ab-initio machine learning models.

I have worked on the Quantum-HP platform, which is dedicated to explicitly incorporating nuclear quantum effects (NQEs) in MD simulations. Currently, my focus lies in integrating Quantum-HP with Tinker-HP's recently introduced Deep-HP machine learning potentials and the Lambda-ABF framework, enabling sub-kcal accuracy of free energies of various molecular systems. An aspect of my work is to understand how NQEs impact the free energy of both molecular and biomolecular systems within the newly developed Q-AMOEBA water force field. 

A significant portion of my research has been dedicated to the development of enhanced sampling techniques, with a special emphasis on incorporating machine learning algorithms. The main focus has been on two objectives: improving conformational sampling and enhancing the accuracy of free energy calculations.

Among the key algorithms I have developed are the adaptive sampling algorithm, Gaussian-accelerated Molecular Dynamics in dual-water mode, and Hamiltonian Reservoir Replica Exchange Molecular Dynamics. Currently, I am working on Lambda-ABF, a cutting-edge method known for its sampling efficiency and high-precision in estimating free energies.