The conference will feature five keynote presentations, and several invited and contributed talks in two parallel sessions, running from lunchtime on the 11th September (Wednesday) and concluding with lunch on the 13th (Friday), covering topics of Energy Materials and Interfaces, Advances in Quantum Chemistry, Density Functional Theory, Simulations of Soft Matter, Biomolecular Simulations and Machine Learning among others.

A poster session will take place in the evening of Wednesday, with buffet dinner provided.

Keynote talks

Beyond Crystallinity and Throughput: Machine Learning Accelerated Materials Discovery for Energy Conversion and Storage

Prof. Karsten Reuter

Fritz Haber Institute, Germany

More performant and durable materials are urgently needed to further drive the transition to a sustainable energy system. Unfortunately, accelerated materials discovery is in this field presently still more claim than practical reality. Computational screening approaches hinge on efficient descriptors that only reflect nominal materials properties of the crystalline bulk, simple bulk-truncated surfaces or idealized lattice-matching interfaces. They can thus not account for the substantial, complex and continuous structural, compositional and morphological transitions at the working surfaces or interfaces of catalysts, electrolyzers or batteries. Accelerated experimental discovery in turn still suffers from severe throughput limitations, as easily automatable human steps are rarely limiting the overall workflows. In my talk I will illustrate how modern machine learning (ML) approaches help to overcome these challenges. ML surrogate models, in particular in conjunction with agile active

____________

Accelerating the design of Soft Materials using Machine Learning

Prof. Marjolein Dijkstra

Utrecht University, The Netherlands

Predicting the emergent properties of a material from a microscopic description is a scientific challenge. Machine learning and reverse-engineering have opened new paradigms in the understanding and design of materials. However, this approach for the design of soft materials is highly non-trivial. The main difficulty stems from the importance of entropy, and the ubiquity of multi-scale and many-body interactions. In this talk, I will address questions like: Can we use machine learning to autonomously identify local structures, detect phase transitions , classify phases and find the corresponding order parameters, can we identify the kinetic pathways for phase transformations and can we use machine learning to coarse-grain our models ? Finally, I will show in this talk how one can use machine learning to reverse-engineer the particle interactions to stabilize quasicrystals, liquid crystals, and crystals.

____________

Computational Discovery of New Materials for Singlet Fission in the Solid State

Prof. Noa Marom

Carnegie Mellon University, US

Intermolecular singlet fission (SF) is the conversion of a photogenerated singlet exciton into two triplet excitons residing on different molecules. SF has the potential to enhance the conversion efficiency of solar cells by harvesting two charge carriers from one high-energy photon, whose surplus energy would otherwise be lost to heat. The development of commercial SF-augmented modules is hindered by the limited selection of molecular crystals that exhibit intermolecular SF in the solid state. Computational exploration may accelerate the discovery of new SF materials. The GW approximation and Bethe–Salpeter equation (GW+BSE) within the framework of many-body perturbation theory is the current state-of-the-art method for calculating the excited-state properties of molecular crystals with periodic boundary conditions. In this talk, I will discuss the usage of GW+BSE to assess candidate SF materials, as well as its combination with low-cost physical or machine learned models in materials

____________

TBC

Prof. Sarah Harris

University of Sheffield, UK

Tbc

____________

TBC

Prof. Michele Ceriotti

EPFL, Switzerland

Tbc

Invited talks

DE-FF and MACE-OFF: Data-driven interatomic potentials for molecular simulations

Dr Daniel Cole

Newcastle University, UK

Drawing on computational methods that are based around training to extensive condensed phase physical property and quantum mechanical datasets, I will describe some of our efforts to design accurate and transferable inter- and intra-molecular potentials, with a view to applications in condensed phase atomistic modelling and computer-aided drug design.

I will explain how recent collaborations with the Open Force Field Initiative enable the development of a fast, accurate alternative to the Lennard-Jones non-bonded potential [1]. With OpenFF, we developed Smirnoff-plugins as a flexible framework to extend the software stack to include custom force field functional forms. We deployed the infrastructure that OpenFF has provided for optimising parameters against condensed phase data, to train a transferable, small molecule force field based on a double exponential functional form (DE-FF). The automated framework allowed us to train and test a full small molecule force field in just a matter of weeks (as opposed to many years for traditional force fields), with promising accuracy in the condensed phase.

Finally, I will describe MACE-OFF23, a transferable force field for organic molecules created using state-of-the-art machine learning technology and first principles reference data. MACE-OFF23 demonstrates the remarkable capabilities of short-range models by accurately predicting a wide variety of gas and condensed phase properties of molecular systems, including dihedral scans, descriptions of molecular crystals and liquids, and even properties of a solvated small protein.

[1] Horton, J., et al. Digital Discovery, 2023, 2, 1178. https://doi.org/10.1039/D3DD00070B

[2] Kovács, D., Moore, J.H., et al. (2023). arxiv, https://arxiv.org/abs/2312.15211

____________

Probing Structural Subtleties in Anti-Perovskite Solid Electrolytes

Dr Karen Johnston

Durham University, UK

The rechargeable lithium-ion (Li-ion) battery is considered the technology of choice for energy storage in a wide array of portable electronic devices. However, its application is limited by its use of liquid electrolytes, which are known to pose a serious fire and safety risk. All-solid-state Li-ion batteries are attracting considerable attention as possible alternatives to conventional liquid electrolyte-based devices as they present a viable opportunity for increased energy density and safety. In recent years, a number of candidate materials have been explored as possible solid electrolytes, including garnets, Li-stuffed garnets, Li-rich anti-perovskites (LiRAPs) and thio-LISICONs. In particular, LiRAPs, including Li3−xOHxCl, have generated considerable interest, based on their reported ionic conductivities (~10−3 S cm−1).1,2 However, until recently, their lithium and proton transport capabilities as a function of composition were not fully understood.

Current research efforts have focused on the synthesis and structural characterisation of Li3−xOHxCl using a combination of high-resolution powder diffraction and variable-temperature multinuclear solid-state NMR spectroscopy with ab initio molecular dynamics. We will demonstrate that Li-ion transport is highly correlated with the proton and Li-ion vacancy concentrations. In particular, we will show that the Li ions are free to move throughout the structure, whilst the protons are restricted to solely rotation of the OH− groups. Based on these findings, and the strong correlation between long-range Li-ion transport and OH− rotation, we have proposed a new Li-ion hopping mechanism, which suggests that the Li-rich anti-perovskite system is an excellent candidate electrolyte for all-solid-state batteries.3 However, to fully understand the mechanism for conduction, multiple, complementary characterisation techniques are required. We will also present our latest findings regarding the room temperature structure of Li2OHCl, which has been the subject of considerable debate within recent years. We will demonstrate the complexities associated with elucidation of this structure and discuss the influence and impact of these structural findings on the ion conduction capabilities of this class of materials moving forward.

[1] Y. Zhao and L. L. Daemen, J. Am. Chem. Soc., 2012, 134, 15042.

[2] A. Emly, E. Kioupakis and A. Van der Ven, Chem. Mater., 2013, 25, 4663.

[3] J. A. Dawson, T. S. Attari, H. Chen, S. P. Emge, K. E. Johnston and M. S. Islam, Energy Environ. Sci., 2018, 10, 2993.

____________

Exploring the Pt(111)-Electrolyte Interface Under Applied Potentials with Ab Initio Molecular Dynamics

Dr Clotilde Cucinotta

Imperial College London, UK

In this talk, I will discuss some complexities in the simulation of electrified interfaces at the nanoscale, focusing on the impact of applied potentials on their physicochemical properties. My approach is based on the development of highly realistic ab initio molecular dynamics models of charged electrode-electrolyte interfaces under bias. I will discuss recent advancements in modelling the double layer of the electrified Pt(111)-electrolyte interface, particularly in terms of its response to the applied electrode potential. This is achieved through the application of electrode-charging and potential control methodologies developed in my group [1-3]. If time permits, I will discuss how the how insights from molecular electronics can lead to a more sophisticated understanding of electrochemical phenomena.

[1] https://doi.org/10.1016/j.electacta.2021.138875

[2] https://pubs.acs.org/doi/10.1021/acs.jpclett.3c03615

[3] https://www.researchsquare.com/article/rs-3788305/v1

____________

Learning from fundamental surface science for atomic layer deposition – an ab initio endeavour

Prof. Dr Ralf Tonner-Zech

Leipzig University, Germany

Atomic Layer Deposition (ALD) emerges as an important technique in addressing the evolving complexities of microelectronics fabrication, offering the potential of atomically precise material construction. But as every practitioner of the method knows, the reality is much more complex and many physical and chemical effects have to be considered, tested and investigated to arrive at the goal of a targeted materials synthesis. A current forefront challenge is achieving area-selective ALD (AS ALD), aiming for selective material growth on target surfaces, with small molecule inhibitors (SMIs) presenting a promising strategy to prevent deposition on non-growth areas.

The nature of the process requires an in-depth understanding of the underlying surface chemistry as well as a mechanistic understanding of the adsorption and decomposition reactions that can happen in the SMI as well as the ALD process steps. Our latest research shines a light on SMI-based AS ALD, offering insights into experimental results and pushing the boundaries toward predictive computational analysis.1–4

At its core, ALD is governed by surface chemistry principles, allowing extrapolations from fundamental studies on molecular-surface interactions, often performed in more experimentally controlled settings. Extensive research into the reactivity of organic adsorbates on semiconductor surfaces has unearthed key reaction mechanisms, bridging insights reactivity known from molecular chemistry.5,6 For metal surfaces, while a local view on chemical bonding may fall short, the interplay with organic adsorbates is still within reach of computational approaches. Our findings include investigations into the stability of the longest known acene chain on Cu(111)7 and the behavior of non-alternant aromatic compounds on metallic substrates.8

One focus of the work is based around understanding the electronic structure and revealing the underlying chemical driving forces using our energy decomposition method for extended systems (pEDA) which allows for quantitative analysis.[6]

[1] Yarbrough, J. et al. Area-Selective Atomic Layer Deposition of Al2O3 with a Methanesulfonic Acid Inhibitor. Chem. Mater. 35, 5963–5974 (2023).

[2] Yarbrough, J. et al. Tuning Molecular Inhibitors and Aluminum Precursors for the Area-Selective Atomic Layer Deposition of Al2O3. Chem. Mater. 34, 4646–4659 (2022).

[3] Oh, I.-K. et al. Elucidating the Reaction Mechanism of Atomic Layer Deposition of Al2O3 with a Series of Al(CH3 )xCl3–x and Al(CyH2y+1)3 Precursors. J. Am. Chem. Soc. 144, 11757–11766 (2022).

[4] Zoha, S., Pieck, F., Gu, B., Tonner-Zech, R. & Lee, H.-B.-R. Organosulfide Inhibitor Instigated Passivation of Multiple Substrates for Area-Selective Atomic Layer Deposition of HfO2. Chem. Mater. (2024) doi:10.1021/acs.chemmater.3c02525.

[5] Pecher, L., Laref, S., Raupach, M. & Tonner, R. Ethers on Si(001): A Prime Example for the Common Ground between Surface Science and Molecular Organic Chemistry. Angew. Chem. Int. Ed. 56, 15150–15154 (2017).

[6] Pecher, L. & Tonner, R. Deriving bonding concepts for molecules, surfaces, and solids with energy decomposition analysis for extended systems. WIREs Comput. Mol. Sci. 9, e1401 (2019).

[7] Ruan, Z. et al. Synthesis of Tridecacene by Multistep Single-Molecule Manipulation. J. Am. Chem. Soc. 146, 3700–3709 (2024).

[8] Klein, B. P. et al. Molecular Topology and the Surface Chemical Bond: Alternant Versus Nonalternant Aromatic Systems as Functional Structural Elements. Phys. Rev. X 9, 011030 (2019).

____________

Reversible simulation to train classical and machine learning potentials

Dr Joe Greener

MRC Laboratory of Molecular Biology, UK

Molecular mechanics force fields are usually developed manually using quantum mechanical (QM) and condensed phase data, whereas machine learning potentials are typically trained on QM data alone. Ideally there would be automated methods that can reproducibly train potentials with any number of parameters to match a variety of data types. Here I present reversible simulation, a development of differentiable simulation, which calculates the gradients of observables with respect to force field parameters over a molecular simulation. Applications to parameterising molecular mechanics water models with various functional forms and a machine learning materials model will be shown. The approach is compared to existing ensemble reweighting methods and has the advantage of allowing fitting to dynamics data.

____________

Can we learn ‘the best’ Markov Model?

Dr Antonia Mey

University of Edinburgh, UK

Markov state models (MSM) have been a widely used tool for the analysis of conformational dynamics of proteins, including protein folding. As it is a statistical and machine learning (ML) model certain choices must be made in constructing an MSM. These choices, or hyperparameters, are often chosen by expert judgement or maximizing variational scores such as the VAMP-2 score. Modern ML and statistical pipelines often use automatic hyperparameter selection techniques ranging from the simple: choosing the best score from a random selection of hyperparameters to the complex: optimization via e.g., Bayesian optimization. Can we use these out-of-the-box optimization techniques to select ‘the best’ MSMs automatically?

In this talk, I will walk you through how we systematically built and assessed over a million MSMs for protein folding models and why model observables and variational scores should only be used to guide model selection. Furthermore, I will show some recent highlights of how we can use MSM to extract longer-than-millisecond timescale dynamics from around 1 ms aggregated simulation time of a kinase.

____________

Machine learning of Gaussian basis sets for use in molecular applications

Dr Grant Hill

University of Sheffield

Gaussian basis sets have a large influence on both the accuracy and efficiency of molecular electronic structure calculations, but there have been few significant advances in their development over recent years. Recent work in our group has focused on using high-throughput calculations and machine learning to produce novel, molecule-optimised basis sets.

The open source Python package BasisOpt [1] has been used to optimise basis set exponents for molecules (rather than the usual optimisation for atomic energies), while the use of Legendre polynomials [2] reduces the dimensionality of the optimisation problem and avoids variational collapse. As a tool for automating the optimisation of basis sets, additional applications of BasisOpt will also be presented, including generation of auxiliary basis sets for use in density fitting, and reducing large basis sets into more efficient ones.

Analysis of data from the molecule-optimised basis sets leads to the use of machine learning (ML) to predict new basis sets for molecular applications. Results of ML basis sets will be presented and compared to existing alternatives, demonstrating that the approach is effective in producing accurate and efficient basis sets.

[1] Shaw, R. A.; Hill, J. G. J. Chem. Phys. 2023, 159, 044802.

[2] Petersson, G. A.; et al. J. Chem. Phys. 2003, 118, 1101–1109.

____________

Surface-Specific Spectroscopy through Machine Learning

Dr David Wilkins

Queen's University Belfast, UK

In this talk, I show how machine-learning predictions of the energies and forces of atomistic systems, along with predictions of their responses to applied electric fields, can be combined with imaginary time path-integral methods to provide fully quantum-mechanical predictions of surface-sensitive sum-frequency generation (SFG) spectra. These calculations require that the polarizations and polarizabilities of the systems be accurately predicted; the former in particular requires some thought about how to build models for bulk polarizations. For this, I describe two approaches that render the polarization straightforward to learn.

____________

Accelerating the prediction of large carbon clusters combining structure search and machine-learning interatomic potentials

Dr Carla de Tomas

Imperial College London, UK

From as small as single carbon dimers up to giant fullerenes or amorphous nanometer-sized particles, the large family of carbon nanoclusters holds a complex structural variability that increases with cluster size. Capturing this variability and predicting stable allotropes remains a challenging modelling task, crucial to advance technological applications of these materials. While small cluster sizes are traditionally investigated with first-principles methods, a comprehensive study spanning larger sizes calls for a computationally efficient alternative.

Here, we combine the stochastic ab initio random structure search algorithm (AIRSS) with geometry optimisations based on interatomic potentials to systematically predict the structure of carbon clusters spanning a wide range of sizes. We first test the transferability and predictive capability of seven widely used carbon potentials, including classical (EDIP, ReaxFF, Tersoff, REBO-II, LCBOP-I, AIREBO) and machine-learning (GAP-20) potentials. Results are compared against an analogous cluster dataset generated via AIRSS combined with density functional theory optimizations. The best performing potential, GAP-20, is then employed to predict larger clusters in the nanometre scale, overcoming the computational limits of first-principles approaches. Our complete cluster dataset describes the evolution of topological properties with cluster size, capturing the complex variability of the carbon cluster family. As such, the dataset includes ordered and disordered structures, reproducing well-known clusters, like fullerenes, and predicting novel isomers.