top of page

Research Themes

High-Performance Computing (HPC)
Algorithms for Digital Chemistry

DALL·E 2024-03-03 12_edited.jpg

At the DMT Lab, we are at the forefront of designing cutting-edge algorithms and software tailored for high-performance computing systems, with our team holding three world records in the field.


Our focus is on harnessing the capabilities of state-of-the-art supercomputing technology, including the world's fastest machines, to simulate and accurately predict the physicochemical behaviour of matter at molecular scales and speeds previously unattainable.


Through our pioneering efforts, we deliver software that achieves computational modelling and simulation precision rivalling that of chemico-physical experiments. We use this software to significantly accelerate the rate of innovation and discovery in chemical R&D, providing a more cost-effective, versatile, and fully automatable complement to physical experimentation.

Achieving predictive accuracy requires the adoption of computational quantum molecular mechanics modelling techniques, also known as quantum chemistry (QC) methods. However, the historical application of QC to large-scale molecular systems, such as those involved in the development of novel therapeutic drugs, has been stymied by its prohibitive computational time complexity of ≥ O(N ) and intrinsic parallel efficiency limitations of available methods and software. 


We unlock the full potential of QC by designing novel high-performance algorithms and software implementations that reduce the computational complexity from quintic and beyond to linear, while being jointly tailored to use at near-peak floating point performance up to tens of thousands of GPUs on the world leading exascale supercomputing platforms. This enables us to apply high-fidelity QC modelling to key technological problems, such as the discovery of new drugs and materials, that have been so far impossible to simulate at the quantum mechanical level due to the large size of the underpinning molecular systems.

Current research foci include the algorithmic redesign of a variety classical and quantum computational chemistry methods for low-time-complexity, massively parallel execution on many GPUs. Modelling methods of interest include: Classical and Ab Initio Molecular Dynamics, Density Functional Theory, Perturbation Theory, Coupled-Cluster Theories, Molecular Fragmentation, Implicit Solvent and Cluster Solvent Models, Quantum Conformational Search, Quantum Gradients and Hessians. 

Artificial Intelligence (AI) for 
Digital Chemistry


The DMT Lab integrates Artificial Intelligence (AI) across various tiers of our digital chemistry technology.

In our pursuit of accelerating the drug and material discovery process, we develop novel machine learning algorithms specifically tailored to predict the energetics and dynamics of molecular systems with near quantum mechanical accuracy. A significant challenge in this area lies in the scarcity of high-quality data crucial for training models to achieve the required precision. While our high-speed, massively parallel quantum chemistry software, EXESS, mitigates this limitation to some extent, it is essential to devise robust methodologies that minimize data requirements and computational overhead while ensuring model integrity. To address this, we incorporate appropriate inductive biases into our model design, facilitating robust learning from constrained datasets. An exemplary technique is the utilization of equivariant neural networks (NN), which maintain symmetry properties inherent in molecular systems, thereby enhancing the robustness and accuracy of predictions.

Generative AI techniques are also pivotal in our endeavor to create promising compounds for therapeutic drugs and novel metamaterials. By leveraging algorithms and models, Generative AI generates potential designs and configurations that traditional methods might overlook, thereby accelerating the human-driven discovery process.

Through sophisticated machine learning approaches, Generative AI explores vast solution spaces, facilitating the discovery of optimized compounds and materials tailored for specific therapeutic or functional purposes. Techniques such as generative adversarial networks (GANs) and variational autoencoders (VAEs) enable the generation of diverse and innovative molecular structures or material compositions, simulating and predicting their properties to inform decision-making towards optimized and application-tailored designs.

Another application of AI within our framework is the generation of models for quantitative structure-activity relationship (QSAR) studies, a key area in drug discovery and design. Deep learning techniques, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs), can build accurate QSAR models that predict the biological activity of compounds based on their structural features or complex descriptors. By automatically extracting intricate patterns and relationships from large datasets, these models provide insights into structure-activity relationships, facilitating the rational design of novel drugs with optimized efficacy and safety profiles.

Autonomous Quantum and AI Protocols for Digital Drug and Material Discovery

DALL·E 2024-03-04 11_edited.png

The digital design process for novel therapeutics and materials is a complex, multi-stage endeavour. In a funnel-like progression, candidate chemicals undergo gradual screening using a pipeline of molecular tests and simulations that become increasingly accurate and computationally demanding. Initially, simple and rapid models assess structural requirements or similarity to known molecules with desirable properties, such as established drugs. Candidates that successfully pass these initial tests then undergo various screening procedures, ranging from physics-based and/or AI-driven models to ascertain if a chemical possesses the desired properties, to molecular dynamics simulations aimed at obtaining more comprehensive thermodynamic statistics regarding their properties. 

Subsequently, the most promising compounds undergo further scrutiny, for example incorporating high-precision quantum mechanics, followed by experimental validation in laboratory settings. Post-lab testing, the most effective compounds serve as the foundation for structural optimizations aimed at enhancing their target properties. 

These optimization are traditionally performed by humans (medicinal and synthetic chemists). These optimized compounds are then subjected to a cycle of screening tests, including virtual molecular dynamics simulations and quantum simulations, along with in vitro experiments, to refine their attributes iteratively.

Across each stage of the design process, a diverse array of computational modeling and simulation techniques may be deployed, each potentially with multiple variations. Moreover, the human-driven nature of each stage, coupled with its intricate linkage to preceding and subsequent phases, presents a non-automated and highly complex framework necessitating diverse file formats and scientific expertise. This human element renders the process both time-consuming and financially burdensome, while also introducing the limitations of human knowledge and biases, particularly in creative phases like structural optimization.

At the DMT Lab, our mission is to revolutionize this intricate process by constructing digital protocols that automate and streamline its complexities. To achieve this, we employ modular, compatible, and integrable software at each stage of the design pipeline, enhanced by AI insights and automation. For instance, generative AI facilitates structural optimizations, machine learning models effectively infer properties from quantum or structural data, and automated decision-making guides the integration of stages within the digital discovery process, optimizing efficiency and efficacy. Through our innovative approach, we strive to accelerate scientific advancement and alleviate the burdens of traditional methodologies, ushering in a new era of discovery and innovation.

bottom of page