RAEng Google DeepMind Research Ready Summer Internships 2026

Supervisor(s):

Margherita Battistotti with Julia Handl

Description:

Many optimisation algorithms, such as metaheuristics, rely on internal parameters that strongly affect their performance. For example, an algorithm used to plan delivery routes can produce solutions of very different quality depending on how these parameters are chosen. Unfortunately, the default settings provided with most algorithms are often not optimal.

Automatic Algorithm Configuration (AAC) is a research area that studies how to automatically find good parameter settings for algorithms based on the problems they need to solve. Most existing AAC methods work offline: they analyse a fixed dataset of past problems and select a single parameter configuration that performs well on average. However, in many real-world applications, problems arrive sequentially over time. For example, a logistics company must compute delivery routes every morning, but customers, locations, and constraints change from day to day. We call this a streaming scenario.

This project explores how to make AAC methods more adaptive in such streaming settings. In particular, we want to investigate whether a class of neural networks called autoencoders can automatically learn useful representations of optimization problems (also called instances) without relying on humandesigned numerical features. Human-designed features (for example, the number of customers in a logistics delivery problem) may seem important to us, but they might not actually influence how well an algorithm performs. Instead, performance may depend on more subtle factors that are difficult for humans to identify.

An autoencoder is a neural network that learns to compress input data into a compact, "latent” representation and then reconstruct it. In this project, we extend standard autoencoders by adding a classification head. The aim is to learn a latent space in which problem instances that are best solved by the same algorithm configurations are located close to one another.

The intern will work with synthetic or benchmark optimisation problem instances and performance data from a fixed set of algorithm configurations. The project will involve:

• Training autoencoders with a classification component to learn unbiased instance representations.
• Evaluating whether these representations can be used to predict which configuration will work best on new, unseen problems.
• Exploring whether changes in the incoming data stream (for example, seasonal effects in logistics problems) can be detected automatically by the model, such as when new instances fall outside the learned latent space.
• Investigating simple strategies to retrain or adapt the model when such changes occur.

The project is exploratory and research-oriented, but it is strongly grounded in practical machine learning and optimisation problems that also arise in industry. It is well suited for students interested in Optimisation, Machine Learning and Artificial Intelligence.

Suggested reading (will be provided as needed):

• Schede, Elias, et al. "A survey of methods for automated algorithm configuration." Journal of Artificial Intelligence Research 75 (2022): 425-487.
• Goodfellow, I., Bengio, Y., & Courville, A. Deep Learning, Chapter 14: Autoencoders.
• Luke, S. Essentials of Metaheuristics, Chapter 0–3.

Skills required:

• Python programming.
• Introductory knowledge of machine learning and neural networks.
• Curiosity about optimisation algorithms and their applications.

Career path enabled:

• Graduate research in AI, machine learning, optimisation, computer science.
• Industry roles such as data scientist or machine learning engineer.

Supervisor(s):

Rebecca Bowler with Omar Rivasplata

Description:

At the cutting edge of astrophysics research is the study of the formation of the first generation of stars and galaxies. Distant galaxies are extremely rare, and because of their long light travel time (billions of years) they are extremely ‘faint’ and challenging to identify in images (e.g. Bowler+20). In current data, we are searching for tens of sources within millions. ‘Contaminants’ that have similar properties vastly outnumber the sources of interest. Imminent arrival of new large datasets (>10,000 times the area, trillions of sources) make the current approach infeasible, and new techniques must be developed to discover patterns in the vast data to understand this fundamental cosmic epoch.

Deep learning models typically needs hundreds of thousands to millions of known training examples, which is impossible in the search for rare sources. Instead, in this project we will test a specialist development of the Zoobot galaxy foundation model (Walmsley+22) to solve this problem. Zoobot implements a deep supervised representation learning pipeline to produce meaningful semantic representations of galaxies, which can be used on new tasks on which the models were not trained. Essential to the success of the project will be an optimised fine-tuning of the model to accommodate dataset shift in the multi-source data, followed by the design and implementation of project specific downstream task(s) to identify truly distant galaxies while minimising the false positive rate due to contaminants (e.g. Walmsley & Scaife 23).

In summary, this interdisciplinary summer internship will develop novel Machine Learning techniques and applying these to cutting-edge astronomy imaging data to search for distant galaxies. The student will become experienced in relevant AI for image analysis, and develop a bespoke implementation (Perez-Ortiz+21) of the Zoobot foundation models (Walmsley+22) to the first data preview from the Vera Rubin Observatory.

Project plan:

• Weeks 1-2 – Literature review of AI/ML approaches to image analysis. Preliminary analysis of available LSST Data Preview and processing of the data into appropriate format.
• Weeks 2-6 – Application of Zoobot (or another appropriate model) to data. Optimisation and testing of the model, using previously identified distant galaxies as test cases.
• Weeks 6-8 – Interpretation of resulting samples of galaxies. Quantification of efficacy through computation of contamination and completeness. Discussion of limitations and possible extensions.

References:

• Walmsley, M et al. 2022, MNRAS, 513, 1581
• Bowler, R et al. 2020, MNRAS, 493, 2059
• Pérez-Ortiz, M et al. 2021, JMLR, 22, 227

Skills required:

• Experience in python programming
• Strong physics and mathematics background
• Desire to learn about machine learning
• Interest in astrophysics

Career path enabled:

• Graduate study
• Industry/ data science

Supervisor(s):

Jiaoyan Chen with Richard Allmendinger

Description:

Goldfire by Accuris is an AI-powered cognitive search platform built for engineering and R&D teams. It can read large volumes of unstructured documents such as internal files, standards, patents, technical reports and web sources, and answer natural language questions with semantic retrieval and extraction, rather than simple keyword matches. Goldfire Lenses, also known as Concept Lanes, are curated views that organise results by facets such as definitions, parameters/values, causes & effects, standards/codes, entities, and more, so users can see the most relevant facts and context around a topic.

The main objective of this internship project is to transform Accuris Goldfire results, especially its Concept Lenses, into a machine-readable graph, such that further development can (1) apply graph-theoretic analytics, such as the discovery of centrality, paths, communities, to reveal critical entities, gaps, and relationships in known knowledge about large, unstructured engineering corpora, and (2) implement uncertain reasoning using technologies like probabilistic graph model, graph representation learning and more recently large language model over the graph for discovering new knowledge, such as hidden relationships between entities and potential facets of the knowledge, so as to improving the experience of knowledge navigation and retrieval.

To this end, the project will use Knowledge Graph (KG), a graph-based knowledge representation method that has been widely used by many organisations and companies (e.g., Google builds its KG for augmenting its search engine; Amazon manages its products as a KG). KG enables flexible knowledge extension and integration, efficient and expressive querying supported by many modern graph databases, graph locality-based retrieval, human friendly visualisation, as well as more expressive conceptual knowledge representation and reasoning via straightforward plugin of ontologies (e.g., RDFS and OWL) and logical rules (e.g., Datalog and Horn).

For implementation, this project will include the following two main work packages:

• KG construction from Goldfire Lenses.
  • This includes (1) ingest and query, where queries will be run on documents, for capturing lens outputs such as definitions, parameters, cause and effect, standards and codes; (2) entity and relation extraction, where the schema is designed, lens items are mapped to the schema with entities like concept, component, standard, clause, parameter and material, and relations between concept and definition, cause and effect, etc.; (3) normalisation and enrich, where entities and relations will be canonicalized and annotated with provenance; (4) storage and service, which will make the KG accessible via query endpoint and APIs, supported by graph database and hybrid indexes.
• Use cases:
  • Based on the constructed KG, the project will explore graph-theorical analytics and KG uncertain reasoning for the following use cases:
    • (1) standards compliance graph for clauses with highest betweenness and shortest paths from a clause to design artifacts;
    • (2) root-cause & mitigation navigator for computing communities of failures and ranking common mitigations;
    • (3) innovation landscaping for revealing emerging concept clusters and bridging concepts with high betweenness.

In the end, the project is to deliver a KG of Goldfire Lenses with accessing APIs and an endpoint, as well as demonstration of the three use cases.

Skills required:

Mandatory:

• Programming and software engineering - especially Python
• Experience and basic skills of data processing
• Foundations of NLP, database and machine learning

Desirable:

• Graph analysis
• Knowledge graph construction, knowledge graph reasoning, information extraction, foundation of recent AI technologies such as large language model.

Partners:

Industrial Partner:

Prof Nawal Prinja, Prinja and Partners Limited

Technology Partner providing Goldfire Software License:

Accuris, large international, HQ based in USA, formerly known as S&P Global Engineering Solutions, is a global technology company that provides engineering intelligence and workflow solutions.

Career path enabled:

• Data or knowledge engineer
• Applied AI or machine learning engineer
• Research engineer
• Enterprise search and knowledge management specialist

Supervisor(s):

Lucas Codeiro with Mustafa Mustafa

Description:

Software written in Python is widely deployed, but it often comes with (security) vulnerabilities that developers must fix. This project proposes integrating Large Language Models (LLMs) with Bounded Model Checking (BMC) to ensure the security and safety of Python code. Large language models will receive intermediate analysis results, along with other verification details, to break down the original verification problem into more manageable sub-problems. We will design a chain-of-thought prompting technique for interacting with LLMs in automated reasoning to increase confidence in output results and optimize program verification. We will use the ESBMC-AI framework as a prototype, which connects Gemini from Google and ESBMC to support the verification and repair of Python code.

The project will run for 8 weeks, divided into three tasks: 
(1) enhance ESBMC-AI’s repair module for Python; 
(2) use LLMs to optimize BMC verification; 
(3) benchmark verification accuracy, scalability, and efficiency.

Deliverables include an improved ESBMC-AI tool, finetuned models, benchmarks, and a public report to enable other organisations to widely use this framework.

This work targets source code verification and repair, especially for Python software [1].

ESBMC has found practical use in industry, including deployment by companies such as Intel and ARM for verifying real-world applications. Its recent LLM extension, ESBMC-AI, has been extensively evaluated on 1,000+ synthetic programs [2]. In other independent work, ESBMC has been combined with LLMs for invariant detection and applied to SV-COMP benchmarks of varying complexity (see [3). The proposed approach could be applied to prove the safety and security of Python code running at Google.

ESBMC is an open-source, permissively licensed, context-bounded model checker based on satisfiability modulo theories for the verification of single- and multi-threaded programs. It does not require the user to annotate the programs with pre- or postconditions, but allows the user to state additional properties using assert-statements, which are then checked as well. Regarding its LLM extension, the BMC component validates or refutes the correctness of the code generated by LLMs.

ESBMC uses the Apache 2.0 BSD 4-clause license and includes third-party code under the same terms. ESBMC-AI is developed under the GNU Affero General Public License v3.0.

ESBMC on GitHub 
ESBMC-AI on GitHub 

Work plan:

Week 1: Review the literature, set up the environment, and get to know the ESBMC framework and how it integrates with LLMs.
Week 2: Build an AI code dataset with keras2c or onnx2c and use the Google Gemini to create vulnerable code samples.
Week 3: Complete dataset creation, begin finetuning the Google Gemini LLM engine.
Weeks 4 and 5: Keep finetuning the Gemini LLM and update ESBMC architecture to support code repair.
Week 6: Connect the finetuned LLM to the ESBMC-AI framework and set up the repair loop.
Week 7: Expand the framework to include memory safety checks and equivalence verification.
Week 8: Comprehensive evaluation, prepare technical report, and GitHub release.

[1] Bruno Farias, Rafael Menezes, Eddie B. de Lima Filho, Youcheng Sun, Lucas C. Cordeiro: ESBMC-Python: A Bounded Model Checker for Python Programs. ISSTA 2024.
[2] Charalambous, Yiannis, Norbert Tihanyi, Ridhi Jain, Youcheng Sun, Mohamed Amine Ferrag, and Lucas C. Cordeiro. "A New Era in Software Security: Towards Self-Healing Software via Large Language Models and Formal Verification." AST 2025.
[3] Wu, Haoze, Clark Barrett, and Nina Narodytska. "Lemur: Integrating Large Language Models in Automated Program Verification." ICLR 2024.

Skills required:

• Background on bounded model checking and large language models
• C++ programming
• Python programming skills

Supervisor(s):

Marie Farrell with Hazel Taylor

Description:

Formal methods allow us to robustly verify that software-driven systems meet their requirements. To do this, requirements are formulated as logical statements for verification. However, during requirements elicitation, requirements are typically expressed as natural-language statements which are not readily amenable to formal verification. NASA's Formal Requirements Elicitation Tool (FRET) helps to bridge this gap by providing a way to automatically translate natural-language requirements into formal, logical properties [1].

There has been much research on specifying requirements for autonomous and cyber-physical systems such as robotic platforms like the Mars rovers, rovers for navigating nuclear environments, wild-fire monitoring and rescue drones, etc. [2, 3] This research tends to focus solely on the robotic system(s) at hand and often omits the fact that there is always a human involved (e.g. as a controller, supervisor or even as a teammate). As these systems become more sophisticated, so do the prevalence of human-robot teams - where humans and robots must work collaboratively to achieve a common goal. In critical settings, there is a need to verify the behavior of such teams, but this starts by eliciting and formalizing their requirements.

In this project, we seek to answer the following research questions:

Q1: Is an existing taxonomy of human-robot teamwork requirements practical in a large-scale use case?

Q2: Can such requirements be formalized using FRET? If not, how can FRET be extended to accommodate such requirements?

To answer these questions, we will work alongside our industry partners (Amentum, NASA Ames Research Center, and Space Solar) to explore the formalization of requirements for a human-robot teamwork scenario. This scenario will likely be one of (1) robot-assisted dressing, (2) assembly of large structures in space, or (3) wildfire inspection and firefighting using drones. The project will comprise the following:

Week 1: Understanding and training on the use of FRET and the current human-robot team requirements taxonomy.
Week 2: Defining the exact dimensions of the case study along with our industry partners.
Week 3: Elicitation of natural language requirements for the use case.
Week 4: Formalization of the elicited natural language requirements using FRET and an associated formal analysis of these requirements using simulator and realizability checker within FRET.
Week 5: A report describing the requirements, their characteristics and proposed extensions to both the taxonomy and FRET strengthen its capabilities to express requirements for human-robot teams.
Week 6: Decide which extensions to implement to the taxonomy and/or FRET and make these changes.
Week 7: Continue to implement the proposed changes and draft a paper on the research results.
Week 8: Write-up.

[1] Giannakopoulou et al. "Formal requirements elicitation with FRET." International Working Conference on Requirements Engineering: Foundation for Software Quality. 2020.
[2] Luckcuck et al. "Formal specification and verification of autonomous robotic systems: A survey." ACM Computing Surveys (2019).
[3] Bourbouh et al. "Integrating formal verification and assurance: an inspection rover case study." NASA Formal Methods Symposium. 2021.

Skills required:

Required:

Strong results in engineering modules and coursework

Desirable:

Experience with formal methods and requirement engineering

Career path enabled:

• This project will support a career path in either academia or industry.
• Project can be expanded into a final year project.
• Writing experience for final year thesis.

Supervisor:

Ali Hassanzadeh

Description:

Professional basketball games are fundamentally driven by interactions between players rather than isolated individual actions. When a specific lineup is on the court, players form a temporary "interaction network" through passing, spacing, and coordinated decision-making. Some lineups consistently outperform expectations, suggesting that chemistry and style play an important role beyond individual talent. However, these concepts are difficult to measure using traditional statistics.

This project aims to develop an AI-based approach to learn latent representations of basketball lineups using graph neural networks (GNNs) applied to play-by-play passing data from professional basketball games (e.g. the NBA). Each lineup will be represented as a graph, where players are nodes and passing interactions form edges. Player-level features (such as role or position) and lineup-level context (such as game state) will be incorporated into the graph representation.

The core research question is the following: can we learn a compact embedding of a lineup that captures its chemistry, playing style, and effectiveness?

The intern will explore modern graph-based representation learning methods, such as graph autoencoders or contrastive GNNs, to learn embeddings where: i) lineups with similar behaviour and outcomes are close together in the learned space, and ii) lineups with very different effectiveness or styles are far apart. These learned embeddings will then be evaluated on downstream tasks, such as predicting: a) points scored per possession, and b) turnover probability.

The intern will also explore clustering the learned lineup embeddings to identify groups of lineups with similar styles, as well as "outlier" lineups that are unusually effective or ineffective relative to their peers.

The project is deliberately exploratory and research-focused. The goal is not to build a production system, but to investigate whether advanced AI models can uncover meaningful structure in complex team interactions. The results will provide insights into how collective behaviour emerges in team sports and will form a reusable methodological foundation for future research on team performance, fairness, and scheduling in sports analytics.

Work plan:

Week 1: Background reading on basketball analytics, graph neural networks, and representation learning; environment setup; familiarisation with the play-by-play dataset.
Week 2: Data cleaning and preprocessing; construction of passing graphs for on-court lineups; basic exploratory analysis.
Week 3: Implementation of baseline models and initial graph neural network architectures; testing simple graph representations.
Week 4: Development of graph-based representation learning models (e.g. graph autoencoders or contrastive GNNs) to learn lineup embeddings.
Week 5: Evaluation of learned embeddings on downstream tasks such as predicting points per possession and turnover probability.
Week 6: Analysis and visualisation of lineup embeddings; clustering and comparison of similar and dissimilar lineups.
Week 7: Model refinement and exploratory extensions (e.g. comparing teams or lineup styles).
Week 8: Preparation of final report and presentation summarising methods, results, and future research directions.

Skills required:

Essential:

• Programming experience in Python
• Familiarity with machine learning

Desirable: 

• Interest in graph-based models or deep learning
• Experience with Pytorch or similar frameworks
• Curiosity about data-drive research

Career path enabled:

• Research oriented careers in AI and Data Science (areas involving graph neural networks, representation learning, data driven decision-making)
• MSc or PhD
• Industry research roles

Supervisor(s):

Dmitry Kangin with Omar Rivasplata

Description:

Designing organic materials that exhibit desirable properties—such as high binding affinities, well-defined structures, and specific functions—is a challenging and critical problem in organic chemistry. Recent generative models like GFlowNet demonstrate strong potential in computational generation. Building on recent theoretical and empirical developments in studying GFlowNets (E.Bengio et al, 2021) as well as the follow-up work Silva et al (2021), we propose to
train a new model which could, in the subsequent stages of research, pave the way towards transforming the precision of organic material sampling beyond the current state-of-the-art. To this end, we will conduct an empirical evaluation using ChemGymRL dataset (Beeler et al, 2023), which will help pave the way towards further theoretical and empirical studies as well as applications in design of organic materials, breaking with the conventional approaches in the field.

GFlowNets models learn probability distributions and representations on latent spaces that abstract essential features of the data, leading to generators capable of synthesising new samples that resemble the distribution in the data’s native high-dimensional space, often conditioned on user prompts in the form of text, target properties or partially-known data. GFlowNet works by optimising the rewards for generating samples of data, such as molecules. In our implementation of GFlowNets, we will use this library on GitHub.

The goal of ChemGymRL is to simulate real-world chemistry experiments, while keeping it simple enough to incorporate into modern machine-learning pipelines. The environment produces the rewards based on the procedure and outcomes of actions taken by the agents. The aim is for ChemGymRL to help bridge the gap between autonomous laboratories and digital chemistry by creating the environment suitable for designing and benchmarking machine learning algorithms. We will use ChemGymRL to simulate the downstream rewards for training GFlowNets. In our implementation, we will use ChemGymRL implementation.

During the project duration, we propose the following timeline:

Week 1: Literature review, familiarising with GFLowNets and ChemGymRL environment, environment setup
Week 2: Implementing the data pipeline with GFlowNets, producing simulation with the baseline implementation
Week 3-4: Introducing ChemGymRL into the GFlowNet pipeline, producing simulations
Week 5-6: Furtner improving the model, producing plots
Week 5-7: Prepare the final report which would outline the experimental findings

• E. Bengio et al (2021) Flow Network based Generative Models for Non-Iterative Diverse Candidate Generation, NeurIPS 2021
• Beeler et al (2023) ChemGymRL: An Interactive Framework for Reinforcement Learning for Digital Chemistry, NeurIPS2023-AI4Science
• Silva, T., Souza, A. H., Rivasplata, O., Garg, V., Kaski, S., and Mesquita, D. (2025). Generalizationand distributed learning of GFlownets. In The Thirteenth
  International Conference on LearningRepresentations.

Skills required:

• Python programming
• Generative Models
• Background in mathematics

Career path enabled:

• Applied or fundamental machine learning.

Supervisor(s):

Andrew Leach with Lia Sotorrios

Description:

The last few years have seen an explosion of AI tools aimed at helping us to discover new medicines. These include Nobel prize winning contributions from Alphafold as well as many others with their origins in either the tech or pharmaceutical industries or academia. While these tools are now being used routinely, it is clear that many of them still suggest molecules that are not possible or sensible. Tools to filter through the output and identify molecules that could actually exist and which might have desirable properties are desperately needed. For about a decade, we have been developing computational tools that allow us to use quantum mechanics to predict how molecules, such as those the AI tools suggest, might interact with proteins; a key step in the activity of most medicines. Quantum mechanics is the fundamental physics that governs molecules and an unintended but important consequence of using it is that any molecule that we study is effectively cross-checked to make sure that it can exist. Our efforts in the last three years have refined our workflow to allow us to process molecules much faster than previously and so we find ourselves perfectly placed to pipe the output from AI tools in and produce a refined list of the best molecules for scientists to make.

In this project, you will run a number of the most prominent AI tools for suggesting new molecules including REINVENT (developed by Astrazeneca) and DeepChem. You will take the molecules that are proposed and apply the Leach group’s theoceptor approach to them. The first step of the theoceptor approach involves applying quantum mechanics to the molecule itself; although this sounds complex, this is a straightforward calculation. Any molecules that fall apart during this step do so because according to quantum mechanics they do not exist. These can therefore be removed from consideration. In the final step, you will compute how the molecules bind to the protein of interest and this will allow the most exciting molecules to be picked out because those are the ones that are likely to have the biggest effect for patients.

The project will provide an opportunity to visit our longstanding collaborators at Medchemica Limited (a company founded by Dr Leach) where you can see how these and other computational tools are used across the pharmaceutical industry and to spend time with professional software engineers. The project is ideally suited for those interested in careers in the pharmaceutical industry or in applied AI development either in industry or academia. It will be highly beneficial but not strictly necessary to have a good understanding of chemical structures and chemistry.

Work plan:

Week 1: Meet the team in Manchester and at Medchemica. Familiarise yourself with the environment and begin using the operating systems and software.
Week 2: Start generating outputs using the REINVENT and Deepchem software.
Week 3: Install and begin using the theoceptor software.
Week 4-6: Generate sets of molecule structures via the AI tools aligned with a number of live group projects, likely to include kinase and ion channel inhibitors (important targets for cancer, cardiovascular and mental health conditions). In parallel, these will be passed forward to quantum mechanical and theoceptor calculations to verify that they are plausible and value-adding molecules.
Week 7-8: Summarising and statistical analysis to assess the success rate of AI tools for suggesting molecules of potential value in drug discovery. Report findings to the team in Manchester and at Medchemica.

Partner:

Medchemica Limited

Skills required:

Understanding of chemistry and chemical structures

Career path enabled:

Industry or academia

Supervisor(s):

Rahul Singh Maharjan with Angelo Cangelosi

Description:

This 8-week research project investigates the use of vision-language models (VLMs) for social signal processing, with a focus on benchmarking their performance for human–robot interaction (HRI) applications. Social cues such as gaze direction, facial expressions, body posture, and interaction-oriented gestures play a central role in human communication. However, reliably interpreting these signals in unconstrained, real-world environments remains a significant challenge for robotic systems. Traditional approaches typically rely on task-specific perception tightly coupled to particular cues or scenarios. Recent advances in VLMs offer a promising alternative for modeling social understanding in a unified, flexible manner.

The primary objective of this project is to systematically evaluate and compare the performance of several state-of-the-art, “open-source vision-language models”—such as LLaVA and LLaMA 3.2 Vision—on datasets containing social interaction scenarios. The student will focus exclusively on open models to ensure reproducibility and accessibility. Benchmarking will be conducted using established social and HRI datasets, including the MuMMER Dataset, the Social-IQ Dataset, and the AVDIAR Dataset, which contain diverse examples of human–human and human–robot interactions.

As part of the project, the student will define one or two well-scoped social inference tasks suitable for systematic evaluation. Example tasks include detecting intent to interact (e.g., whether a person is inviting engagement), estimating social attention (such as gaze direction or attentional focus toward the robot), or recognizing coarse affective states (e.g., positive, neutral, or negative affect). These tasks will be formulated in a way that is compatible with vision-language model inference. Using the selected datasets, the student will design consistent evaluation protocols, including task definitions, prompts, and metrics, to enable fair and meaningful comparison across models. The benchmarking will primarily focus on zero-shot and few-shot learning settings. Model performance will be evaluated using quantitative social signal processing metrics (including accuracy, F1-score), complemented by qualitative analysis to identify typical strengths, limitations, and failure modes, such as missed social cues or hallucinated interpretations.

In the final phase of the project, the benchmarking results will be connected to a simple robotic perception or decision-making scenario. Outputs from the evaluated models, such as predicted engagement or social intent, will be mapped to high-level robot behaviors in a minimal simulation or prototype pipeline. This stage will clearly present how benchmarking outcomes can inform model selection and design choices for socially aware robots. The project will conclude with a research report summarizing the key methodology and experimental results.

Partners:

Dr. Luca Raggioli, PRISCA Lab, University of Naples 'Federico II'.

The project partner will help the student conduct a human-robot interaction experiment on the robot at PRISCA Lab, Napoli, Italy, and CoRoLab, Manchester, UK. (This will depend on the progress of the student.) 

Skills required:

• Strong Python programming; experience with PyTorch
• Strong foundations in computer vision (image and/or video) and NLP
• Familiarity with HuggingFace transformers libraries
• Comfort working with large-scale datasets
• Basic understanding of robotics concepts (perception pipeline, ROS preferred but not required)
• Interested in reading research papers, experiment tracking, and clear writing

Career path enabled:

• Postgraduate study in multimodal learning, or social signal processing, or human-robot interaction
• Applied AI roles involving computer vision + NLP, and embodied AI
• Robotics perception engineer

Supervisor(s):

Christoforos Moutafis with Daniel Burrow

Description:

In this internship, AI and machine learning are used as research tools to accelerate materials characterisation within an existing spintronics programme, rather than as an end in themselves, in the context of an AI-in-the-loop lab.

Magnetic multilayers can host nanoscale spin textures such as stripe domains and skyrmions, which are of strong interest for emerging neuromorphic and probabilistic computing paradigms. The properties and stability of these textures are governed by a small number of underlying magnetic effects. Accurately extracting the material parameters that govern these effects is essential for designing functional devices, but conventional experimental techniques and the subsequent data analysis are often slow, indirect, and resource-intensive.

This is where placing AI in the loop for materials characterisation acceleration can have an impact. This project builds on recent research demonstrating that deep learning models can infer multiple magnetic interaction parameters directly from images of magnetic domain configurations, with a combination of synthetic and experimental data. By training convolutional neural networks on simulated magnetic domain images, the inverse problem can be addressed: predicting physical material parameters from experimentally accessible, binarised microscopy images. This approach offers a high- throughput complement to traditional measurements and aligns with our broader goal of using AI methods to accelerate spintronic materials design for computing applications.

The intern will work on developing and exploring this physics informed machine learning pipeline. Activities will include analysing magnetic domain images, learning how domain morphology encodes physical information, training and validating neural network models, and assessing model robustness when applied to noisy or experimentally realistic data. Particular emphasis will be placed on understanding limitations, such as performance near the boundaries of parameter space and sensitivity to image preprocessing choices.

Rather than focusing on generic AI benchmarking, the project will maintain a strong connection to physical interpretation. Key questions include which geometric features of domain patterns are most informative, how learned relationships reflect known micromagnetic constraints, and how the model could be extended to handle more realistic material effects relevant to experiments. Depending on progress, there may be scope to explore multi-modal fusion of different experimental data.

The project supports both industry and graduate research career pathways. Interns will gain experience at the interface of physics, materials science, and applied machine learning, developing transferable skills in Python programming, data analysis, and scientific modelling. The work is particularly well suited to students interested in AI for physical sciences, neuromorphic hardware, or data driven approaches to experimental research, while remaining grounded in a real materials and devices research environment. This project is hosted in a spintronics and neuromorphic computing research group, not a standalone artificial intelligence lab. Our core expertise lies in magnetic materials, nanoscale devices, and brain inspired hardware.

Work plan:

Week 1: Literature review and introduction to the overall workflow. Read key papers on micromagnetics, convolutional neural networks (CNNs), U-Nets, and other relevant  neural networks. Familiarisation with the scientific context and goals of the project.

Week 2: Familiarisation with tools and initial implementation. Introduction to the primary software and frameworks used in the project (e.g. PyTorch, MuMax3). Train and evaluate an existing U-Net model using a dataset of simulated and experimentally derived magnetic domain images.

Week 3–4: Introduction of a physics-based approach. Process and analyse magnetic domain images, link domain morphology to underlying magnetic parameters, and curate a physics-informed training dataset. Train a supervised U-Net model to learn mappings between magnetisation textures and underlying physics.

Week 5–6: Model refinement and evaluation. Continue experimenting with the physics-informed model, perform inference on unseen and experimentally realistic data, and systematically assess performance. 

Week 7: Generalisation and robustness testing. Explore how well the trained models generalise to data acquired under different experimental conditions or parameter regimes. Synthesis and analysis of results.

Week 8: Consolidation and communication. Prepare final results, figures, and documentation (GitHub). Produce a short technical document and presentation summarising methods, results, limitations, and future directions.

Partners:

The student will be working as part of the NeuroSky team (skyrmionics.org) and the Nanoengineering and Spintronics Laboratory group at the Department of Computer Science, The University of Manchester.

Skills required:

Essential skills, knowledge, and experience:

• Familiarity with coding/Python for data processing and scripting.
• Data analysis/mathematical analysis skills (ability to handle, clean, and preprocess datasets).
• Prior experience with machine learning frameworks (e.g., PyTorch, TensorFlow).
• Interest in machine learning and AI, particularly in how they apply to scientific and technical domains.
• Interest or experience in neural networks such as CNNs.

Desirable skills, knowledge, and experience:

• General interest in nanotechnology or materials science or physics.
• Good verbal communication skills to explain technical outcomes to varied audiences.

Career path enabled:

• Graduate research pathways in physics, materials science, computer science, electrical engineering, or AI for physical sciences
• Cross-disciplinary careers at the interface of experimental science, numerical simulation, and machine learning, suitable for research labs, national facilities, and deep-tech companies.
• Applied AI roles focused on scientific modelling, simulation acceleratio
• PhD and MSc preparation for students interested in neuromorphic computing, computational materials science, or data-driven approaches to experimental physics and device engineering.
• Industrial R&D roles in advanced materials, semiconductor technology, spintronic devices, and hardware-aware AI, where hybrid physics–machine learning workflows are increasingly used.