DAY 4

Computational drug discovery has become an increasingly important tool in the search for new drugs to treat a wide range of diseases. One of the critical challenges in this field is simulating the complex behavior of biomolecules at the atomic level, which can be computationally very expensive and time-consuming. The thermodynamics and kinetics of drug-target binding play a crucial role in understanding and optimizing the interactions between drugs and their target molecules. To address this challenge, molecular dynamics (MD) and related approaches combined with machine learning and artificial intelligence play a crucial role in accelerating the exploration of the conformational space of molecules. These methods also enable researchers to compute the free energy differences between different states of a molecule and estimate thermodynamic and kinetics properties such as binding free energies and residence time. This talk overviews various enhanced sampling methods combined with machine learning in computational drug discovery, as well as their advantages and limitations. The talk focuses on using these methods for free energy and kinetics estimations, reporting on the significant limitations toward accurate estimations for large datasets of compounds. The talk then describes recent applications in drug discovery and discusses the challenges and limitations of these methodologies when applied to biological systems of increasing complexity. In conclusion, the talk illustrates how MD, enhanced sampling methods, and machine learning can improve the efficiency and effectiveness of computational drug discovery and accelerate the development of new therapeutics to treat significant unmet medical needs in the era of precision medicine.

One of the most exciting developments in recent years is the development of graph deep learning interatomic potentials such as the Materials 3-Body Graph Network (M3GNet), that have near-universal coverage across the entire periodic table. Such universal interatomic potentials (UIPs) have broad applications in the dynamic simulations and discovery of materials. However, current UIPs are still hampered by the lack of accurate potential energy surface (PES) data. In this talk, I will discuss these limitations and how advanced sampling approaches based M3GNet latent structural features can be used to generate high-quality PES data in an efficient manner. In addition, I will show that these approaches can also be used to quantify the uncertainty of a UIP for a structure. I will conclude with some observations on the relative performance of invariant versus equivariant UIP architectures, and a perspective on the challenges involved in UIPs with additional physics such as charges and magnetism.

The European Spallation Source (ESS) aims to become the world's leading neutron source. Emphasizing the crucial role of scientific computing from its inception, ESS prioritizes software, hardware, simulations, and data management to increase the impact from research with neutron scattering at ESS. This presentation introduces ESS and its data pipeline, highlighting the potential for the electronic structure simulation community to benefit from and contribute to ESS's scientific endeavors. Over recent decades, the integration of electronic structure simulations with neutron scattering has grown steadily, enhancing experimental data interpretation. Utilizing ray-tracing simulation programs like McStas allows for precise comparison between simulated and experimental data in a like-to-like manner, facilitating a 'digital twin' approach that includes experimental artifacts. The presentation will feature exemples and address associated challenges.

In this talk I will discuss techniques and workflows targeted at the simulation of molecular dynamics protocols based on ab initio calculations, machine-learned (ML) potentials of ab initio quality and including quantum nature of the nuclei. I will discuss aspects of electronic-structure code efficiency in HPC architectures to generate data in order to train ML models and recent improvements in the i-PI code which decrease communication overhead when running over many HPC nodes, allowing to reach nanoseconds of molecular dynamics simulation per day with ML potentials.

Understanding the dynamic processes at solid-liquid interfaces in electrochemical devices like batteries is key to developing more efficient and durable technologies for the green transition. Fundamental and performance-limiting interfacial processes like the formation of the Solid-Electrolyte Interphase (SEI) [1] spans numerous time- and length scales. Despite decades of research, the fundamental understanding of structure-property relations remains elusive. Ab initio molecular dynamics (AIMD) generally provides sufficient accuracy to describe chemical reactions and the making and breaking of chemical bonds at these interfaces [3]. Still, the cost is prohibitively high to reach sufficiently long time- and length scales to ensure proper statistical sampling [4]. Machine learning (ML) potentials offer a potential solution to this challenge. Still, training ML-based potentials capable of handling activated processes in organic or aqueous electrolytes remains a fundamental challenge since the potential must capture both intra- and intermolecular interactions in the electrolyte and during chemical reactions at the interface [4]. Here, we present new approaches using foundation models [5], graph neural networks [6] and new transition state training sets [7] for chemical reaction networks, and machine/deep learning models to predict the spatio-temporal evolution of electrochemical interphases [3]. We also discuss the development of active learning methods to accelerate segmentation of microstructures in non-destructive 3D imaging techniques, such as X-ray nano-holo-tomography, and enable the visualization of battery electrodes. Finally, we discuss how such models trained on multi-sourced and multi-fidelity data from multiscale computer simulations, operando characterization, high-throughput synthesis, and testing, to provide uncertainty-aware and explainable ML for early prediction of patterns from chemical spectra (Figure 1) [8].

The Ada Lovelace Centre (ALC) is a center of expertise for scientific computing within STFC with the primary objective of maximizing the scientific impact of the STFC large scale national facilities, Diamond, ISIS and the Central Laser Facility.

The ALC is seeking to cultivate integrated approaches to address and solve the scientific data challenges experienced by facilities and their users. In early discussions with the facilities, the integration of theory with experiments, the provision of theory as a service, and linking users to modelling and simulation experts were highlighted as priorities. Here we will discuss the ALC and efforts to shape a program to meet the modelling and simulation needs of the facilities.

In this talk, I will present the method for developing sustainable software tools, and engaging user communities, that is currently being implemented by members of the Scientific Computing Department (SCD) of the Rutherford Appleton National Laboratory, at the Science and Technologies Facilities Council of the UK.

At present, the method comprises two strands: a) development of theoretical models and sustainable tools for the interpretation of muon experiments; and b) management of the computational workflows associated with the processing, and interpretation, of muon and x-ray absorption spectroscopy (XAS) experiments performed at the lab.