Back to overview

Next-Generation Multiscale Molecular Dynamics: Promoting Computational Chemistry with Artificial Intelligence

English title Next-Generation Multiscale Molecular Dynamics: Promoting Computational Chemistry with Artificial Intelligence
Applicant Roethlisberger Ursula
Number 185092
Funding scheme Project funding
Research institution Laboratoire de chimie et biochimie computationnelles EPFL - SB - ISIC - LCBC
Institution of higher education EPF Lausanne - EPFL
Main discipline Physical Chemistry
Start/End 01.05.2019 - 30.04.2023
Approved amount 951'144.00
Show all

Keywords (10)

Biomimetic Design; multiscale modelling; QM/MM ; Photovoltaics; Computational Chemistry; machine learning; molecular dynamics; Computational Biophysics; Protein Engineering; artificial intelligence

Lay Summary (German)

Next-Generation Multiscale Molecular Dynamics: Promoting Computational Chemistry with Artificial Intelligence
Lay summary

Anwendungen künstlicher Intelligenz (KI) revolutionieren unseren Alltag, sei es durch Sprach- oder Gesichtserkennung, selbstfahrende Fahrzeuge, Verkehrsüberwachung, intelligente Roboter, Suchmaschinen und Video-Streaming oder durch spezialisierte Anwendungen wie die Voraussage von Konsumentenvorlieben oder in der medizinischen Diagnostik. Ein ähnlicher Quantensprung findet momentan im wissenschaftlichen Rechnen statt, wo zunehmend maschinelles Lernen (ML) angewandt wird. KI eröffnet der computergestützten Chemie (CC) neue Möglichkeiten, sei es für die Berechnung der elektronischen Struktur oder für molekulare Simulationen. Dies ergibt sich aus den immer grösseren Datenmengen, die durch die Verwendung immer effizienterer Computer und Algorithmen anfallen, und die die Anwendung von ML-Methoden einerseits bedingen und andererseits überhaupt möglich machen. KI besteht jedoch nicht nur aus ML, sondern umfasst auch andere Anwendungen wie die Datenanalyse durch Bayesische Inferenz oder die Optimisierung durch evolutionäre Algorithmen (EA).

Mit diesem Forschungsantrag wollen wir das fruchtbare Zusammenspiel zwischen KI-Algorithmen und traditionellen CC-Anwendungen in enge Rückkoppel. Dies ist eine logische Erweiterung der Forschungsarbeit, die im Rahmen unserer zwei vorherigen SNF-Projekte durchgeführt wurde, in denen wir ML-Algorithmen wie z.B. Eigenschaftsselektion und kausale Inferenz angewandt haben, um hochdimensionale Simulationsdaten zu analysieren, und in deren Rahmen wir die EA-Toolbox EVOLVE entwickelt haben, die eine Reihe von Anwendungen im Bereich des inversen Designs von molekularen Materialien und Proteinen ermöglicht. Hier sollen diese Methoden erweitert werden, indem wir sie mit Entwicklungen und Anwendungen in unserem langjährigen Forschungsgebiet der quantenbasierten Multiskalen-Molekulardynamik im Grund- und angeregten Zustand kombinieren.

Direct link to Lay Summary Last update: 18.04.2019

Responsible applicant and co-applicants


Associated projects

Number Title Start Funding scheme
165863 Next Generation First-Principles Based Molecular Dynamics with Application to Biomimetic and Materials Design 01.05.2016 Project funding
157107 Modulation of the site specificity of binding of metal-based drugs to chromatin 01.10.2014 Project funding


Artificial Intelligence (AI) is revolutionizing our daily life from speech and facial recognition to self-driving cars, traffic cameras, intelligent robots, search engines, and video streaming to expert systems to predict consumer preferences and perform medical diagnosis. A similar quantum jump is currently starting to take place in computational sciences where an increasing number of applications are starting to emerge that apply machine learning (ML). AI also opens new possibilities for Computational Chemistry (CC) both in electronic structure calculations and molecular simulations. This is a natural evolution since the ever-increasing amount of data (generated by more powerful computers and more efficient algorithms) necessitates and enables the use of ML techniques. However, AI is not limited to ML but also includes other approaches such as analysis tools based on Bayesian inference or evolutionary algorithm (EA) based optimization.In this proposal, we want to promote the fruitful interplay of AI algorithms with traditional CC approaches via close feedback loops (CC ? AI: CC data used for AI, AI predictions used in turn for advancing CC applications). This is a natural extension of our work in the framework of our two previous SNF grants where we introduced ML algorithms such as feature selection and causality inference to analyze high-dimensional simulation data and developed the EA toolbox EVOLVE for a broad range of inverse design applications (in molecular materials and protein design). Here we want to further extend these approaches and combine them with our long-standing efforts in the development and application of quantum based multiscale Molecular Dynamics (MD) simulations in ground and excited states. In particular, we propose the development of a completely new highly flexible multiscale interface to the CPMD ( code called MIMIC. MIMIC is based on a multiple program-multiple data strategy and will allow for flexible combinations of different electronic structure and force fields while maintaining fast communication between the processes. MIMIC will thus allow multi-shell MM/QML/QMH schemes (with different parts of the system represented by different methods) as well as multiple- time step (MTS) based multi-rung QML-QMH approaches (in which a lower level QML method is used to drive and accelerate the actual dynamics with a higher level QMH method) that we have developed during the last two SNF grants. Furthermore, we want to enhance the capabilities of MIMIC through extensive combinations with AI approaches. In particular, we will use genetic algorithms (GA) to explore (and push) the accuracy limits of current exchange-correlation (XC) functionals in Density Functional Theory (DFT) via a GA-DFT approach. We will also develop a new GA based approach (GA-CI) to reduce the cost of high-level wavefunction based electronic structure methods (Full Configuration Interaction (FCI) and multireference active space methods CASCF and CASPT2) by several orders of magnitude. We also want to extend our ML based MD to a fully adaptive ML-MD approach with learning on-the-fly that can also be used as lower level method in MTS based combinations. This powerful new simulation methodologies can be applied to a large range of applications, here we will focus on three areas: the de-sign of biomimetic catalysts, the development of metal based two-centre anticancer drugs and the study of metal halide perovskites for photovoltaic applications.