The Schmidt AI in Science Postdoctoral Fellowship, a programme of Schmidt Sciences, aims to accelerate the next scientific revolution by supporting talented postdoctoral researchers to apply AI techniques across the natural sciences, engineering and mathematical sciences. This initiative adds to Schmidt Sciences’ existing philanthropic efforts to support the development and application of AI in innovative ways.
The programme is cross-disciplinary and spans the full breadth of MPLS Division, which gives it the ability to bring together different parts of the AI landscape in novel ways. Another key element of the AI in Science programme is the training aspect, providing Fellows with all they need as non-AI specialists to use AI in their respective research fields. In addition, the Eric and Wendy Schmidt AI in Science Postdoctoral Fellows have the opportunity of becoming Associate Research Fellows at the University’s newest College: Reuben College. Founded in 2019, Reuben College aspires to create a community of scholars embracing opportunities for interdisciplinary collaboration, and developing initiatives to generate wider impacts and positive benefits from research, entrepreneurship and public engagement.
Professor Jim Naismith, Head of Division for the Mathematical, Physical and Life Sciences, said: “Oxford stands at the forefront of AI research, and witnessing the transformative impact of the Schmidt AI in Science program on researchers within the MPLS Division is truly inspiring. It's remarkable to see how AI is envisioned to address a vast array of challenges, igniting imaginations and propelling innovation forward. We extend our heartfelt gratitude to Schmidt Sciences for their invaluable support and vision in making this pioneering program possible.”
The projects of the new Schmidt AI in Science Fellows are summarised below.
Micah Bowles, Physics: ‘Robust multimodal galaxy embeddings’
To understand the evolution of galaxies, astronomers study their characteristic features (e.g. spiral arms). Machine learning has been used successfully to retrieve certain properties from millions of galaxies! To retrieve other properties, new models must be trained which can be costly and time-consuming. This project will research the application of a new class of AI models trained on different observations of the same objects, giving the resulting model the ability to identify a broader range of characteristics.
Jialun Chen, Engineering Science: ‘A machine learning based method for wave modelling in coastal areas’
Accurate wave forecasting is crucial for various applications, such as optimising vessel route planning or designing robust and economical coastal structures. Numerical wave models are used for understanding ocean wave conditions, but they have limitations in coastal environments owing to factors such as complex shoreline and topography. Machine learning has the potential to achieve a precise representation of coastal wave dynamics, and could substantially reduce the computational costs.
Daniel Dehtyriov, Maths: ‘Physics-based Deep Learning Models for Turbulent Flows with Applications to Renewable Energy’
Fluid flow simulation is integral to modelling various physical systems, but the computational demands associated with turbulence modelling remain a challenge in applications such as wind turbine simulations. Improvements in modelling can hence greatly enhance the economic feasibility of wind-energy projects. The proposal aims to develop a machine learning-based model which will provide a low-computational cost, high-accuracy simulation method for improved turbine design and increased renewable power generation.
Lydia France, Biology: ‘Signatures of Motion: AI-Driven Insight into Dynamic Wing Morphing in Birds’
The project aims to use AI to gain insight into wing and tail movement in flight, developing general methods that can be applied to other birds, in order to inform bioinspired designs and to study natural motion more broadly. Rough path theory will be used to find signatures in the motion of wings in flight, and apply machine learning algorithms which can uncover hidden information and find underlying physics in complex dynamical systems.
Amy Hinsley, Biology: ‘Using AI to develop predictive ‘nowcasting’ of the illegal and unsustainable wildlife trade’
Efforts to address the illegal wildlife trade are hampered by poor-quality and incomplete data, and a lack of predictive understanding of where and when shifts in species, products or trade methods may occur. Applying AI techniques would enable predictive insights into emerging wildlife trade threats, in particular by using ‘nowcasting’, which produces near-term predictions, and is designed for cases when lag-times in data availability hinder effective decision-making.
Veronika Juraskova, Chemistry: ‘Different flavours of machine learning potentials for modelling processes in complex solvated environments’
Solvents are integral components of chemical processes. Most commonly used solvents are toxic to humans and sources of pollution and ideally need replacing with environmentally friendly alternatives. Computational modelling offers insights into chemical reactions at the molecular level and may guide the design of novel synthetic paths. However, modelling chemical reactions in sustainable solvents remains challenging and computationally demanding. This research aims to develop machine learning-based potentials, offering a fast and accurate approach to modelling chemical reactions in complex solutions.
Milan Klöwer, Physics: ‘Climate predictions of precipitation probabilities with online learning’
Climate models provide crucial information for climate change mitigation and adaptation, but they also lack accuracy with respect to many atmospheric variables. Correctly predicting precipitation changes is enormously important, as the associated extreme events (droughts and floods) have large socio-economic impacts. This project aims to build a physics-based atmospheric model that also learns automatically from decades of observational precipitation data, demonstrating how online machine learning can improve simulated precipitation.
Henry Lloyd-Laney, Computer Science: ‘MLMLLM: Machine Learning, Multiple Likelihoods and the Laws of Metallo-proteins’
The proposed research aims to advance the field of bio-electrochemistry by leveraging advanced machine learning techniques to gain a deeper understanding of electron-transfer reactions in complex systems. This has the potential to drive the development of novel catalysts and biosensing technologies with significant real-world applications in sustainable hydrogen production and disease detection. It showcases a pioneering approach to bio-electrochemistry, combining expertise in electrochemistry, statistical inference, Bayesian statistics, and machine learning.
Laura Mansfield, Physics: ‘Developing trustworthy machine learning parameterisations with uncertainty quantification for climate models’
In climate science, a key area of interest is the application of machine learning to subgrid-scale modelling or ‘parameterisations’. Traditional physics-based parameterisations often rely on assumptions and can introduce errors and computational bottlenecks. Machine learning approaches have been explored, but these introduce a source of uncertainty that must be thoroughly understood. This research aims to quantify uncertainties arising from multiple sources with the ultimate aim of revolutionising the next generation of climate models.
Brian Moser, Physics: ‘Probing the Higgs Potential with Graph Neural Networks’
Measuring the Higgs boson’s properties is of great importance for particle physics. Its most crucial unobserved property is that it should interact with itself, which involves searching for the incredibly rare case of a single Higgs boson splitting into two Higgs bosons. Graph Neural Networks will be used for the first time to separate this small signal from a multitude of background processes, paving the way towards measuring the Higgs boson self-interaction.
Cait Newport, Biology: ‘Lost landmarks: Using AI to investigate the impacts of turbidity on visually-guided navigation behaviour of wild reef fish’
Understanding how animals move in their natural environments is key to understanding the basis of many behaviours. One key technique, video photogrammetry, is challenging to use in underwater environments. This research will develop new computer vision tools to enable 3D reconstruction of aquatic animal movements and their environments. Resulting tools will transform how biologists measure and monitor the movements and behaviour of wild aquatic animals.
Jinzhao Sun, Physics: ‘Learning ground-state properties of quantum materials using neural network states’
The interesting behaviours exhibited by new quantum materials are difficult to simulate accurately using conventional methods because of their complex interactions. Neural network representation of quantum states is promising in solving quantum problems. This project will apply quantum-inspired neural networks and quantum machine learning to efficiently capture the complex correlations in materials, which will then be applied to find the most stable configuration of quantum materials.
Jan Christoph Thiele, Chemistry: ‘From noise to knowledge: deep learning for weak signal analysis in ultra-sensitive microscopy and beyond’
In biochemical and pharmaceutical research, understanding interactions between proteins is of paramount importance. Mass Photometry (MP) has emerged as a promising microscopy technique for studying these interactions at the single-molecule level, but faces some limitations. The project aims to expand MP's applicability, building on existing deep learning approaches, including leading methods for super-resolution microscopy and specialised AI tools.
Samvida Venkatesh, Statistics: ‘Multi-pronged approach to systematically identify and validate the transcriptional effects of non-coding genetic variation’
Up to 85% of genetic variants associated with human traits are in the ‘non-coding’ genome and alter gene expression levels rather than protein structure. It is challenging to decipher the precise effect of these variants. This project will benchmark statistical and machine learning models for linking ‘non-coding’ genetic variants to changes in gene expression and transcription. Using a combined model may reduce experimental costs, and improve the ability of models to capture long-range genome interactions.
Mengyun Wang, Materials: ‘Intelligent quantum meta-optics enabled by machine learning for the generation and shaping of non-classical light’
Quantum light sources, which produce photonic quantum bits, have historically relied on cumbersome laser systems, making them difficult to use in the field. Meta-optics (two-dimensional arrays of nanostructures or ‘meta-atoms’), offer an unprecedented platform for versatile quantum light sources and promise huge potential for quantum photonics technology. This project aims to use cutting-edge machine learning techniques to assist the design of multi-functional metasurfaces in quantum-optics applications.
Further information