Periodic Reporting for period 1 - PHYMOL (Physics, Accuracy and Machine Learning: Towards the next-generation of Molecular Potentials)
Periodo di rendicontazione: 2023-02-01 al 2025-01-31
PHYMOL trains researchers to develop accurate physical models, advance theoretical tools, and integrate them with machine learning to create robust molecular potentials for complex applications.
The PHYMOL programme is international (through secondments abroad) and interdisciplinary (spanning physics, chemistry, biology). We have strong industrial involvement, and aim to train students aware of commercial potential dreaming in fundamentals of molecular science and prepares Doctoral Candidates to meet future challenges in both academia and industry.
At the core of PHYMOL are weak intermolecular interactions, which drive essential processes such as protein folding, charge transfer in solar cells, and collisions in interstellar gas clouds. Understanding these forces is crucial for improving and controlling such phenomena. Modelling has become a key tool in interpreting experiments and guiding materials design, but accurately describing weak interactions remains a major challenge for many theoretical approaches.
Despite the importance of this field, there is currently no comprehensive training that combines theoretical foundations, method development, rigorous validation against experiment, and model building using both physical insight and machine learning. PHYMOL addresses this gap by uniting leading experts in chemical physics, molecular simulation, and machine learning, along with academic and industrial partners .The programme uses both deep physical understanding and modern machine learning to build improved tools for simulating how molecules behave and interact.
The PHYMOL programme is international (through secondments abroad) and interdisciplinary (spanning physics, chemistry, biology). We have strong industrial involvement, and aim to train students aware of commercial potential dreaming in fundamentals of molecular science and prepares Doctoral Candidates to meet future challenges in both academia and industry.
At the core of PHYMOL are weak intermolecular interactions, which drive essential processes such as protein folding, charge transfer in solar cells, and collisions in interstellar gas clouds. Understanding these forces is crucial for improving and controlling such phenomena. Modelling has become a key tool in interpreting experiments and guiding materials design, but accurately describing weak interactions remains a major challenge for many theoretical approaches.
Despite the importance of this field, there is currently no comprehensive training that combines theoretical foundations, method development, rigorous validation against experiment, and model building using both physical insight and machine learning. PHYMOL addresses this gap by uniting leading experts in chemical physics, molecular simulation, and machine learning, along with academic and industrial partners .The programme uses both deep physical understanding and modern machine learning to build improved tools for simulating how molecules behave and interact.
The PHYMOL project comprises 11 interconnected research projects, each contributing to the overall scientific objectives of the network. The main technical and scientific achievements to date are as follows:
1. Towards the accurate description of the induction energy in SAPT.
We derived new method for dispersion free (induction energy) in variational manner using second-quantization equations. The method was implemented within the Psi4 code and is currently undergoing testing in the repulsive regions of intermolecular interactions.
2. Intermolecular interactions in excited states:
Potential energy surfaces were computed for He–CO, Li–CO, N2, and CO–N2 systems. Bound and resonant states were identified, and cross-section calculations were performed. SAPT-based methods were applied to gain insight into the nature of these excited-state interactions.
3. Collision-induced absorption and high-resolution spectra calculated from ab initio potentials.
Potential energy and induced dipole surfaces were generated for noble gas atom–atom interactions. A custom code was developed to compute collision-induced absorption spectra. The resulting data will be submitted to the HITRAN database.
4. Reparametrisation of semiempirical models.
We developed the Functional Group Corrections approach to enhance semiempirical methods, implemented it within MOPAC, and benchmarked its performance against machine-learning-based methods achieving very good results.
5. Development of intermolecular force-fields with many-body dispersion interactions.
The project has not started yet as recruitment is still pending.
6. Consistent treatment of polarization and charge-delocalization in many-body systems.
We are currently formulating the method using density matrices and have produced a comprehensive manuscript detailing the formalism in first quantization based on density matrices. Implementation is underway in Psi4 code.
7. How intermolecular interactions shape polymorphic energy landscapes.
We explored machine learning force fields trained on small-cell polymorphs for pre-relaxation of larger structures. While pre-trained models offered modest improvements, results were comparable to those from classical force fields. Tests were conducted on polymorphs of benchmark molecular crystals.
8. State-of-the art modelling of new quantum materials: surface-supported metal atomic quantum clusters (AQC)
We investigated AQCs and their interactions with benzene and coronene to model graphene-supported clusters. Quantum effects such as Jahn–Teller distortions and conical intersections are also being examined. Existing methods are being refined to better capture AQC interactions.
9. Implicit machine-learning solvent models for confined spaces.
We successfully designed and implemented a free energy model. The next stage focuses on generating a benchmark dataset to validate the model against experimental data.
10.Adapting state-of-the-art modelling of new quantum materials to industry-standard opensource big data analysis tools.
We developed a workflow to predict fluorescence quantum yields (FLQY) using machine learning, based on DFT excited-state calculations in ORCA. The protocol includes solvent shell generation and energy decomposition via the GKS-EDA scheme.
11.Density-mapped FFs: Rapid prototyping of force-fields based on physical and ML mappings onto the electronic density.
We developed evaluated methods for extracting atomic properties from molecular density and trained graph neural networks (GNNs) for conformationally flexible molecules. The resulting models yield distributed multipoles for use in molecular dynamics simulations with Tinker.
1. Towards the accurate description of the induction energy in SAPT.
We derived new method for dispersion free (induction energy) in variational manner using second-quantization equations. The method was implemented within the Psi4 code and is currently undergoing testing in the repulsive regions of intermolecular interactions.
2. Intermolecular interactions in excited states:
Potential energy surfaces were computed for He–CO, Li–CO, N2, and CO–N2 systems. Bound and resonant states were identified, and cross-section calculations were performed. SAPT-based methods were applied to gain insight into the nature of these excited-state interactions.
3. Collision-induced absorption and high-resolution spectra calculated from ab initio potentials.
Potential energy and induced dipole surfaces were generated for noble gas atom–atom interactions. A custom code was developed to compute collision-induced absorption spectra. The resulting data will be submitted to the HITRAN database.
4. Reparametrisation of semiempirical models.
We developed the Functional Group Corrections approach to enhance semiempirical methods, implemented it within MOPAC, and benchmarked its performance against machine-learning-based methods achieving very good results.
5. Development of intermolecular force-fields with many-body dispersion interactions.
The project has not started yet as recruitment is still pending.
6. Consistent treatment of polarization and charge-delocalization in many-body systems.
We are currently formulating the method using density matrices and have produced a comprehensive manuscript detailing the formalism in first quantization based on density matrices. Implementation is underway in Psi4 code.
7. How intermolecular interactions shape polymorphic energy landscapes.
We explored machine learning force fields trained on small-cell polymorphs for pre-relaxation of larger structures. While pre-trained models offered modest improvements, results were comparable to those from classical force fields. Tests were conducted on polymorphs of benchmark molecular crystals.
8. State-of-the art modelling of new quantum materials: surface-supported metal atomic quantum clusters (AQC)
We investigated AQCs and their interactions with benzene and coronene to model graphene-supported clusters. Quantum effects such as Jahn–Teller distortions and conical intersections are also being examined. Existing methods are being refined to better capture AQC interactions.
9. Implicit machine-learning solvent models for confined spaces.
We successfully designed and implemented a free energy model. The next stage focuses on generating a benchmark dataset to validate the model against experimental data.
10.Adapting state-of-the-art modelling of new quantum materials to industry-standard opensource big data analysis tools.
We developed a workflow to predict fluorescence quantum yields (FLQY) using machine learning, based on DFT excited-state calculations in ORCA. The protocol includes solvent shell generation and energy decomposition via the GKS-EDA scheme.
11.Density-mapped FFs: Rapid prototyping of force-fields based on physical and ML mappings onto the electronic density.
We developed evaluated methods for extracting atomic properties from molecular density and trained graph neural networks (GNNs) for conformationally flexible molecules. The resulting models yield distributed multipoles for use in molecular dynamics simulations with Tinker.
Central to our approach is the derivation and implementation of a unique, dispersion-free interaction energy. This is achieved through a variational principle using an ansatz inspired by symmetry-adapted perturbation theory (SAPT). The method yields robust results even for complexes that challenge standard perturbation-based approaches. Early benchmarks show improved accuracy and stability, highlighting its potential impact on high-level molecular modelling. A related approach is being pursued in parallel, introducing a general model for charge delocalisation in non-covalent interactions.
The project also produced new datasets for atmospheric and astrophysical communities. We computed collision-induced absorption coefficients for mixed noble gas pairs across a wide temperature range. These results are complemented by highly accurate potential energy surfaces for key atmospheric species, specifically CO and N2. Together, these advances offer more reliable input for atmospheric modelling and remote sensing. Further exciting results are expected as the project progresses.
The project also produced new datasets for atmospheric and astrophysical communities. We computed collision-induced absorption coefficients for mixed noble gas pairs across a wide temperature range. These results are complemented by highly accurate potential energy surfaces for key atmospheric species, specifically CO and N2. Together, these advances offer more reliable input for atmospheric modelling and remote sensing. Further exciting results are expected as the project progresses.