JGI support – Jean Golding Institute News

AI in Health Awardees 2025-2026

Posted on 19 November 202521 November 2025 by kerry.turcsi

Over summer, the Elizabeth Blackwell (EBI) and Jean Golding (JGI) Institutes together with the Faculty of Health and Life Sciences strategic research support fund, and University Hospitals Bristol and Weston NHS Foundation Trust, ran a pump-priming funding call to support innovative applications of AI in health or biomedical research. This funding call came with the expectation that the funded activities would provide a basis for developing and submitting external bids for future research programmes and projects that use or address AI in health and biomedical research contexts.

We are excited to announce 10 projects involving more than 30 researchers supported by this funding. Check out the successful awardees and their projects below.

AI-assisted personalisation of neurostimulation

Petra Fischer, School of Physiology, Pharmacology and Neuroscience
Conor Houghton, School of Engineering Mathematics and Technology

*Left to right: Petra Fischer, and Conor Houghton*

Dystonia is a heterogenous neurological disorder, which causes involuntary muscle contractions, often resulting in pain and severely restricted movement, affecting millions of people worldwide.

A key challenge in neuroscience is understanding how brain networks process sensory input to control movement. Neural synchronisation plays a vital role in organising this activity, occurring both locally and across distant regions. Excessive synchronisation is linked to disorders like dystonia, Parkinson’s, and schizophrenia, and targeted modulation has emerged as a promising therapy.

Dr Fischer’s lab uses non-invasive, phase-specific vibrotactile stimulation to selectively enhance or disrupt synchronisation in dystonia with the aim to improve symptoms. Currently it is still unclear whether local or interregional modulation drives therapeutic effects. The interdisciplinary team will use existing dystonia data and AI-assisted analysis to:

Map brain-wide effects of stimulation, and
Predict outcomes and effective stimulation parameters based on neural data to replace a trial-and-error based search procedure

Findings will support an MRC funding application to develop a clinical stimulation tool, with potential extension to other techniques like transcranial electrical stimulation for direct cortical targeting.

AI-Organoid: A Smart Predictive Platform for Advanced Neurological Modelling

James Armstrong, Bristol Medical School
Qiang Liu, Engineering Mathematics & Technology

*Left to right: James Armstrong and Qiang Liu*

This interdisciplinary project brings together biomedicine and AI to develop AI-Organoid, a predictive, interpretable platform for tracking and forecasting the development of organoids—lab-grown cell models that mimic human organs. Organoids are vital for reducing animal testing and studying human-specific diseases, but their inconsistent growth limits pharmaceutical applications.

Building on promising pilot data (74% accuracy in predicting brain organoid outcomes), the project will refine AI-Organoid to improve reproducibility and provide mechanistic insights into organoid development. Beyond forecasting organoid outcomes, the platform will identify when developmental trajectories diverge and provide mechanistic insights into the underlying biology of brain organoid growth.

The EBI-JGI grant will support data collection, model training, and dissemination, enabling future applications in disease modelling (e.g., neurodevelopmental disorders) and expansion to other organoid types (e.g., ovarian, intestinal, liver). Outputs will include publications, open-source tools, and a foundation for commercialisation and further funding bids.

An AI-integrated lung-on-a-chip platform for the rapid screening and optimisation of mesenchymal stem cell secretome therapeutics

Wael Kafienah, School of Biochemistry and Cellular and Molecular Medicine
Lucia Marucci, School of Engineering Mathematics and Technology
Darryl Hill, School of Biochemistry and Cellular and Molecular Medicine

*Left to right: Wael Kafienah, Lucia Marucci, and Darryl Hill*

Inflammatory lung diseases, such as acute respiratory distress syndrome (ARDS), are devastating conditions with high mortality and no effective drug treatments. A promising new therapy involves using the cocktail of healing molecules secreted by mesenchymal stem cells (MSCs). These cells can be manipulated to optimise the secretome composition towards a specific therapeutic target. However, identifying the most potent secretome composition is a major bottleneck, relying on slow, laborious methods and animal models that poorly predict human responses.

This project aims to develop an AI-integrated Lung-on-a-Chip (LoC) platform to accelerate discovery of effective MSC therapies for inflammatory lung diseases like ARDS, which currently lack treatments. By mimicking lung inflammation and analysing cell responses in real time, the AI will identify optimal MSC secretome compositions more efficiently than current methods.

The pilot will deliver proof-of-concept data, including a high-accuracy AI model and a rich imaging and gene expression dataset, laying the foundation for reducing reliance on animal models and enabling rapid development of regenerative therapies with commercial potential.

An Integrated AI and machine learning platform to enable high throughput, precision oncology driven drug testing

Deepali Pal, School of Biochemistry and Cellular and Molecular Medicine
Colin Campbell, Mathematics, Engineering and Technology
Stephen Cross, Wolfson Bioimaging Unit, Faculty of Life Sciences
Rihuan Ke, School of Mathematics

*Left to right: Deepali Pal, Colin Campbell, Stephen Cross, and Rihuan Ke*

250 children in the UK die from cancer each year. Leukaemia, affecting the human bone marrow, is the commonest children’s cancer. Yet it is a rare disease, which makes studying new treatments in clinical trials challenging. Therefore preclinical prioritisation is key, requiring predictive patient-derived models. However, hospital samples are difficult to cultivate, and where complex tissue-like biomimetic models to allow patient-sample cultivation have been engineered, these are hard to read, making output data inaccessible.

This project will develop an AI-powered bioimaging analysis tool to accurately detect leukaemia cells within complex bone marrow microenvironments, enabling predictive personalised drug screening.

The integrated machine deep learning and AI tool will analyse 3D bioprinted organoids to distinguish leukaemia from morphologically similar bone marrow cells, overcoming limitations of marker-based identification. The interdisciplinary team will apply image processing, expert annotation, and algorithm development to validate the tool.

Impact includes a proof-of-concept precision oncology organoid platform that generates high-throughput, interpretable drug screening data within clinically relevant timeframes. The project also offers commercial potential in the growing organoid market including applicability across other diseases.

Rational AI Driven Target Acquisition from Genomes (RAIDTAG)

Darryl Hill, School of Cellular and Molecular Medicine
Sean Davis, School of Chemistry

*Left to right: Darryl Hill, and Sean Davis*

This pilot study uses AI to accelerate the discovery of highly repetitive DNA sequences for rapid microorganism identification—critical for healthcare, agriculture, and food production. These ‘repetitive signatures’, absent in closely related species, offer precise, cost-effective diagnostic markers.

Building on proof-of-concept using gold nanoparticle detection, the project will deliver a minimal viable product (MVP): an AI pipeline that automates marker discovery and validates candidates in the lab. This interdisciplinary effort, combining expertise in AI and computer science with nanomaterials and diagnostics, will provide the foundation for future external funding applications and translational research.

Predicting PD-L1 status from H&E slides using AI

Tom Dudding, Bristol Dental School
Qiang Liu, School of Engineering Mathematics and Technology
Sarah Hargreaves, Bristol Dental School
Miranda Pring, Bristol Dental School

*Left to right: Tom Dudding, Qiang Liu, Sarah Hargreaves, and Miranda Pring*

There are approximately 377,700 newly diagnosed mouth cancers each year worldwide. Despite treatment, survival rates remain low, and side effects are life changing. Some people with early-stage cancer unexpectedly experience poor outcomes, such as recurrence or early death, and these high-risk patients are often hard to identify. New ways to detect and manage these high-risk cancers are needed.

PD-L1 is a marker found on many cells, including mouth cancer cells. It helps cancers hide from the immune system and may contribute to poorer outcomes. This marker can however be targeted using drugs like Pembrolizumab, which block PD-L1 and help the patient’s immune system fight the cancer. Because of this, PD-L1 is now an important marker used in clinical practice, to guide treatment decisions for people presenting with late-stage head and neck cancers.

This project aims to develop an AI tool to predict PD-L1 expression in mouth cancer directly from digital histology slides, bypassing costly and limited lab tests.

Using the HN5000 cohort—750 digitised slides with linked biospecimens and long-term follow-up—the pilot will create a proof-of-concept AI model. This will support future funding bids, improve diagnostic equity, and expand access to immunotherapy in both NHS and global settings. Deliverables include a validated AI tool for PD-L1 detection, benchmarking against immunohistochemistry to establish reliability, and preliminary analysis to underpin external bids enabling translation of AI-enabled PD-L1 testing into multi-centre validation and ultimately routine clinical practice.

Genetic doppelgangers: using AI to reveal the true face of streptococcal disease

Alice Halliday, Biochemistry and Cellular & Molecular Medicine
Colin Campbell, Engineering Mathematics and Technology
Rachel Bromell, Biochemistry and Cellular & Molecular Medicine
Anu Goenka, Bristol Medical School
Sion Bayliss, Bristol Veterinary School

*Left to right: Alice Halliday, Colin Campbell, Rachel Bromell, Anu Goenka and Sion Bayliss*

This project aims to use AI tools to evaluate and develop diagnostics to distinguish between Streptococcus pyogenes (GAS) and related bacterium, Streptococcus dysgalactiae subspecies equisimilis (SDSE). GAS and SDSE are very similar genetically, such that they are akin to ‘genetic doppelgangers’. Modern DNA-based tests (qPCR) have not been well evaluated for SDSE detection and struggle with closely-related bacteria. This diagnostic confusion impacts our understanding of responsible pathogens and clinical consequences, with increasing evidence that SDSE’s disease burden is significantly underestimated.

Using genetic material from bacteria isolated from clinical throat swabs, this new interdisciplinary team will build on Bristol AI expertise to develop a machine learning classification tool for distinguishing these bacterial species, based on genome ‘k-mers’.

Combining traditional microbiology, novel DNA-based assays and cutting-edge ML analysis of genome sequence data, the team aim to evaluate and design diagnostic tools that accurately identify both pathogens. This could lead to improved diagnostic capabilities, enhanced disease surveillance, better outbreak investigations, and improved patient outcomes.

Bristol Respiratory Infection Dashboard (BRID Project)

Andrew Dowsey, Bristol Veterinary School
Raul Santos-Rodriguez, Engineering Maths and Technology;
Maha Albur, Consultant Microbiologist
Peter Muir, Consultant Clinical Scientist
Paul North, Business Support Manager and Data Analytics, Severn Pathology, North Bristol NHS Trust and UKHSA
Amy Carson, Academic Clinical Fellow
Gavin Deas, Doctor in Training
Marceli Wac, Engineering Maths and Technology
Jack Stanley, Academic Clinical Fellow, Severn Pathology, North Bristol NHS Trust and University of Bristol

*Left to right: Andrew Dowsey, Raul Santos-Rodriguez, Maha Albur, Peter Muir, and Marceli Wac*

This project will develop the AI-powered Bristol Respiratory Infection Dashboard (BRID) at Severn Pathology, serving the South West region of the UK. By integrating real-time data from pathology and care sources, BRID will enable early detection of respiratory infection trends and support targeted interventions.

Respiratory Tract Infections (RTIs) are a major cause of hospital admissions, with over 400,000 cases in 2024 and winter surges up to 80%. Despite available treatments, disparities in vaccine uptake and care access persist. Real-time surveillance is essential to guide equitable, effective responses and improve outcomes.

This project supports NHS England’s strategy for managing acute respiratory infections (ARIs) through integrated care and digital innovation. Backed by medical directors from both merging trusts, the project will compare AI modelling using local Electronic Patient Record (EPR) data versus the South West Secure Data Environment (SWSDE), evaluating data quality, linkage, and implementation.

A co-designed dashboard with clinicians will guide funding bids to scale the platform, aiming to reduce admissions, optimise resources, and improve public health.

Automated image analysis to facilitate the incorporation of quality assurance measures into surgical RCTs

Natalie Blencowe, Bristol Medical School
Michael Wray, School of Computer Science
Anni King, Bristol Medical School
Sheraz Marker, GOLF study, University of Oxford
Nainika Menon, GOLF study, University of Oxford

*Left to right: Natalie Blencowe, Michael Wray, Anni King, Sheraz Marker, and Nainika Menon*

This project aims to develop an AI model to streamline quality assurance (QA) in surgical randomised controlled trials (RCTs), addressing bias caused by variability in surgical technique and skill. Using annotated videos from the GOLF trial, the AI will assess key operative steps based on anatomical visibility as a proxy for quality.

This project provides valuable pilot work to further the application of AI to surgical videos, enabling QA processes to be efficiently embedded into surgical RCTs, meaning they can be adopted more widely. In turn this will improve RCT quality, ultimately improving patient outcomes. There is also potential for these methods, powered by AI, to be used in ‘real time’ during operations to alert surgeons if a key step has not been fully completed, immediately improving surgical quality. Both these applications have wider implications outside of research studies. In routine clinical practice, they could shorten surgeons’ learning curves through the provision of bespoke, real-time feedback. This could transform the way surgeons learn, as well as optimising patient care.

Explainable AI for early categorisation of child deaths: real-time insights for prevention

Karen Luyt, Bristol Medical School
Edwin Simpson, School of Engineering Mathematics and Technology
Brian Hoy, Bristol Medical School
James Gopsill, School of Electrical, Electronic and Mechanical Engineering
David Odd, School of Medicine, Cardiff University

*Left to right: Karen Luyt, Edwin Simpson. Brian Hoy, James Gopsill, and David Odd*

This project brings National Child Mortality Database analysts from the Faculty of Health and Life Sciences and AI experts from the Faculty of Science and ngineering together to develop an explainable high-confidence early categorisation system for the cause of child deaths.

The Child Mortality Analysis Unit (CMAU) at the University of Bristol is the national hub for analysing statutory child death data in England. CMAU links over 25,000 notifications and 19,500 completed reviews with national datasets across health, education, social care, policing, and child safety. This enables the identification of patterns, causes, and risk factors in child mortality, informing preventative action and national policy.

Now, in partnership with analysts from the National Child Mortality Database and leading AI experts, CMAU is developing a pioneering early categorisation system to identify suspected causes of child deaths in real time. This innovation will enhance national surveillance and accelerate public health responses.

Future work will expand to unstructured data from documents such as clinical notes, unlocking insights and spotting patterns that are currently difficult to detect manually. This will mark a major leap forward in understanding and preventing child deaths.

An explainable, high-confidence early categorisation system could be a game-changer, revolutionising how services across sectors monitor, respond to, and ultimately prevent child mortality.

The project team will be sharing the findings via this website, LinkedIn, academic publications and industry events. Are you a national health provider who would like to do something similar? Please reach out and contact us to learn more.

Telling Tales: Building a Folk Map of St Lucia

Posted on 7 October 20252 December 2025 by kerry.turcsi

JGI Seedcorn Funding Project Blog 2024/25: Leighan Renaud

Screenshot from the prototype. Includes a video of a man and the transcript, interactive buttons at the top and a map that changes location with the video — *Figure 1: A screenshot from the prototype map*

In a research trip funded by the Brigstow Institute, a small research team and I met on the Eastern Caribbean island of St Lucia in the Summer of 2024 and spent 10 days immersed in the island’s folk culture. Our research was guided by the question: how do folk tales live in the 21st century Caribbean? We were particularly interested in learning more about the story of the ‘Ti Bolom’ (a Francophone Creole term that literally translates to ‘little man’): a child-sized spirit, often thought to be servant of the devil, that is summoned into the world to do its master’s nefarious bidding. The story has survived for generations on the island, and although the more specific details of this story shift depending on the storyteller, the moral of the story (which warns against greed) typically remains the same. Working with a local videographer and cultural consultant, we recorded a number of iterations of the Ti Bolom story, and conducted interviews about the island’s folk culture, for archival and analytical purposes.

The stories we heard were captivating. We heard from a diverse collection of people with different experiences of the island, its geography and its storytelling culture. The Ti Bolom stories exhibited a variety of cultural influences, from European Catholicism to West African folklore and beyond. The stories we heard also often made reference to very specific places in St Lucia (including villages and iconic locations) as well as occasionally pointing to connections with other neighbouring islands. This small archive of Ti Bolom stories demonstrated the fluid and embodied nature of folk stories and also suggested that there might be a mappable ‘folk landscape’ of the island.

We were interested in exploring innovative and interactive methods of digitally archiving Caribbean folk stories in such a way that honours their embodied nature, and we were curious about the potential of using folk stories as a decolonial mapping method (‘mapping from below’). JGI Seedcorn funding was secured to test whether we could build a ‘folk map’ of St Lucia. Working closely with Mark McLaren in the Research IT team, the aims of this exploratory project were to:

Investigate existing map-based storytelling approaches
Create folk map prototype(s) to demonstrate potential interfaces and functionalities
Document all findings, keeping in mind the potential for future projects (i.e. mapping stories across multiple islands)

Mark understood our decolonial vision immediately, and took a very considered and meticulous approach to building the prototype map, which features stories and interviews from four of the storytellers we recorded in St Lucia. He asked that I provide transcripts for each video, and a list of locations mentioned. Although St Lucia’s official national language is English, they speak Kwéyòl (a French-based creole language) locally, which means that some of the place names used by storytellers are not necessarily their ‘official’ names. This meant that I needed to be careful about ensuring my translations and transcriptions were correct, and I spoke with storytellers and other contacts in St Lucia to validate some of the locations that feature in the stories and interviews.

Mark helped us to test several levels of interactivity with the map. As such, when one chooses a story to view, the filming location and places mentioned are listed so that a user might click through them at their own pace. Simultaneously, when the video is playing and a place is mentioned, the map automatically moves to the new location. This function demonstrates the existence of the folk geographies we hypothesized during our original research project. Folk stories draw their own maps, demonstrating intricate webs of connections, both within the island and beyond.

There were ethical considerations that arose during the development of the map. Folk stories in the Caribbean are considered true. It is often the case that storytellers reference real people, and locations given can identify those people. This was not necessarily an issue we faced when building the prototype, but it did prompt us to consider how much and what kind of information the map should ethically include were it to be developed.

As an interactive folk-archiving method, mapping folk landscapes has the potential to be an innovative and visually arresting output resource that brings Caribbean folk cultures into dialogue with Digital Humanities, and makes these stories accessible for digital and diasporic audiences. The prototype we developed has proven the concept of a Caribbean folk landscape, and this has been pivotal as we develop a grant application for AHRC funding. Our hope is to secure funding to explore, archive and map three kinds of folk stories told across three islands in the Eastern Caribbean.

Thank you to Mark McLaren from the Research IT team for producing a prototype map.

To find out more about the project, feel free to check out our Instagram page, or contact me via email at Leighan.renaud@bristol.ac.uk

This project was talked about in the fourth episode of Bristol Data Stories. You can listen to it here.

Widening Participation (WP) Research Summer Internships

Posted on 13 August 202513 August 2025 by kerry.turcsi

The Widening Participation (WP) Research Summer Internships provide undergraduates with hands-on experience of research during the summer holidays, with the aim of encouraging a career in research. Interns gain professional experience and knowledge through a funded placement in their chosen subject. This also supports application for postgraduate study and other research jobs.

This year, the JGI was very pleased to support four internships through the WP scheme. Each of the interns has provided valuable support to an array of diverse and interesting projects related to their fields of interest. We are delighted by the feedback that we have received from their project supervisors and look forward to watching their future progress. Read on for more information on their projects and their experience.

Frihah Farooq

My name is Frihah, and I’m a third year undergraduate studying Mathematics here at the University of Bristol. My academic interests centre around applied data science and machine learning, and this summer I worked on a project involving the General Practice Workforce dataset published by NHS Digital. My focus was on building tools that could bring accessibility to data that is often scattered and difficult to navigate.

The aim of the project was to automate the downloading and linkage of open-access datasets, specifically in the context of healthcare services. Many of these records are stored in files with inconsistent formats and structures, often requiring manual effort to piece together a consistent narrative. I developed a codebase in R that could search for the appropriate files, extract the relevant information, and construct a complete dataset that can be used for longitudinal analysis without the need for repeated intervention. While the code was built around the workforce dataset, the methodology generalises well to other datasets published by NHS Digital.

One observation from the final merged dataset was the trend of decreasing row counts, likely due to restructuring, alongside an increase in the number of recorded variables, a sign that data collection has grown more sophisticated in recent years. This experience strengthened my foundation in data automation and my ability to work with evolving and imperfect data; skills I know will benefit me as I move further into research.

If you’d like to get in touch, you can reach me at cc22019@bristol.ac.uk

Grace Gilman

Grace's poster on Fair Tales project — *Grace Gilman’s headshot*

Hello, my name is Grace Gilman and I am starting my third year studying Computer Science with Artificial Intelligence at the University of Bath. I am hoping to go into academia in the future and pursue computing research specifically with medical applications. You can contact me at gcag20@bath.ac.uk.

Over the six weeks I have been participating in a research internship here at the University of Bristol, supported by the Jean Golding Institute. I have been working on a data science project called ‘Using AI to Study Gender in Children’s Books’, for the team Fair Tales, supervised by Chris McWilliams. During my internship I experimented with image analysis using ChatGPT and Vertex AIi, for future integration into the Data Entry app that Fair Tales is producing to semi-automate character and transcript input. I have also been contributing to the database architecture and search and filtering options for users to interact with the database. Some of my work has been analysing the corpus of children’s books using SQL, one pattern I found was that the difference between mother and father characters(1:0.75) is even more pronounced for grandmothers and grandfathers(1:0.5).

During my time at this internship, I have become much more confident in my abilities to work on a project as well as code that will be used in a research setting. I have learnt more of how research is conducted and what skills are needed for this, and become more sure of an academic future.

Imogen Joseph

Research poster on 'R packages to guide handling of missing data' — *Imogen Joseph*

I am currently studying a Neuroscience MSci with a Year in Industry at the University of Bristol. I’m going into my final year, having just completed a placement year in Southampton General Hospital undertaking clinical research in neonatal respiratory physiology. I’m particularly interested in a career in academia and more specifically looking at molecular mechanisms behind disease for drug discovery.

This summer, I helped in the development of an R package, ‘midoc’ (Multiple Imputation DOCtor, found on CRAN), designed to guide researchers in analysis with missing data under my supervisor Elinor Curnow. I created several functions that resulted in the display of a summary table of missing data, alongside optional graphs to visualise the distributions of their missing data. This allows the user to explore what is actually missing, and additionally make inferences on whether missingness is random or related to particular variables.

Before coming into this internship, my R ability was limited to self-teaching via youtube videos. Ample training was provided in this project but more than anything, throwing myself in and actually writing code has been so beneficial to my learning. This knowledge is extremely useful for a career in research – I was even able to apply my acquired skills onto the work carried out in my placement, and used R to analyse the data I gathered.

I am very grateful for this opportunity given to me under the JGI and will take what I’ve learnt with me into whatever I do next!

You can contact Imogen at imogenjoseph26@gmail.com

Sindenyi Bukachi

Using Big Data to Rethink Children’s Rights (bsindenyi@gmail.com)

MSci Psychology and Neuroscience, University of Bristol (Year 3)

Sindenyi Bukachi holding their research poster on 'Investigating attitudes towards children's rights (in education)' — *Sindenyi Bukachi holding their research poster*

Initially, the project was quite open – the only brief was to explore attitudes towards children’s rights using big data. My early research into Reddit threads, news stories and real-world discourse helped narrow our focus to something more urgent and measurable: children’s right to participation, specifically in educational settings as both my supervisors are based in the School of Education. This became the foundation for the rest of the project, and my supervisors later decided to take it forward as a grant proposal.

Over the first few weeks, I learned how to do structured literature reviews using academic databases like ERIC, build Boolean search strings, and track findings across a spreadsheet. I explored how participation is talked about and measured, and the themes I identified – like tokenism, power struggles between adults, and the emotional toll of being “heard” but not actually listened to – became central to our research direction.

In the second half, I moved from qualitative sources to dataset analysis. I used R and RStudio to explore datasets from the UK Data Service. I learned to work with tricky file types (.SAV, .TAB), use new packages, extract variables, visualise trends, and test relationships between predictors — all while thinking critically about how these datasets (often not made for this topic) could reflect participation and children’s agency.

I’ve gained confidence in data science, research strategy, and independent problem-solving – all skills I’ll take forward into my dissertations and future career. I’m so grateful to Dr Katherin Barg, Professor Claire Fox, and the JGI for the support and trust throughout.

MagMap – Accurate Magnetic Characteristic Mapping Using Machine Learning

Posted on 25 June 202527 June 2025 by kerry.turcsi

PGR JGI Seed Corn Funding Project Blog 2023/24: Binyu Cui

Introduction:

Magnetic components, such as inductors, play a crucial role in nearly all power electronics applications and are typically known to be the least efficient components, significantly affecting overall system performance and efficiency. Despite extensive research and analysis on the characteristics of magnetic components, a satisfactory first-principle model for their characterization remains elusive due to the nonlinear mechanisms and complex factors such as geometries and fabrication methods. My current research focuses on the characterization and modelling of magnetic core loss, which is essential for power electronics design. This research has practical applications in areas such as the fast charging of electric vehicles and the design of electric motors.

Traditional modelling methods have relied on empirical equations, such as the Steinmetz equation and the Jiles-Atherton hysteresis model, which require parameters to be curve-fitted in advance. Although these methods have been refined over generations (e.g., MSE and iGSE), they still face practical limitations. In contrast, data-driven techniques, such as machine learning with neural networks, have demonstrated advantages in addressing multivariable nonlinear regression problems.

Thanks to the funding and support from the JGI Institute, the interdisciplinary project “MagMap” has been initiated. This project encompasses testing platform modifications, database setup, and neural network development, advancing the characterization and modelling of magnetic core loss.

Outcome

Previously, a large-signal automated testing platform is produced to evaluate the magnetic characteristics under various conditions. Fig. 1 shows the layout of the hardware section of the testing platform and Fig. 2 shows the user interface of the software that is currently used for the testing. With the help of JGI, I have managed to update the automated procedure of the platform including the point-to-point testing workflow and the large signal inductance characterizing. This testing platform is crucial for generating the practical database for the further machine learning process as its automated function has largely increased the testing efficiency of each operating point (approx 6-8s per data point).

Labelled electrical components in a automated testing platform — *Fig. 1. Layout of the automated testing platform.*

Code instructions for the interface of the automated testing platform — *Fig. 2. User interface of the automated testing platform.*

Utilizing the current database, a Long Short-Term Memory (LSTM) model has been developed to predict core loss directly from the input voltage. The model shows a better performance in deducing the core loss than traditional empirical models such as the improved generalized Steinmetz equation. A screenshot of the code outcome is shown in Fig. 3 and an example result of the model for one material is shown in Figure 4. A feedforward neural network has been tried out as a scalar-to-scalar model to deduce the core loss directly from a series of input scalars including the magnetic

flux density amplitude, frequency and duty cycle. Despite the accuracy of the training process, there are limitations in the input waveform types. Convolutional neural networks have also been tested before using the LSTM as a sequence-to-scalar model. However, the model size is significantly larger than the LSTM with hardly any improvement in accuracy.

Code for the demo outcome of the LSTM — *Fig. 3. Demo outcome of the LSTM.*

Bar chart showing ratio of data points against relative error code loss (%) — *Fig. 4. Model performance against the ratio of validation sets used in the training.*

Future Plan:

Although core loss measurement and modelling is a key issue in industrial applications, the reason behind these difficulties is the non-linear relationship between the magnetic flux density and the magnetic field strength which is also known as the permeability of the magnetic material. The permeability of ferromagnetic is very sensitive to a series of external parameters including temperature, induced current, frequency and input waveform types. With an accurate fitting between the relationship of magnetic flux density and field strength, not only

the core loss can be precisely calculated but also the current modelling method that is used in Ansys and COMSOL can be improved.

Acknowledgement:

I would like to extend my gratitude to JGI for funding this research and for their unwavering support throughout the project. I am also deeply thankful to Dr. Jun Wang for his continuous support. Additionally, I would also like to express my appreciation to Mr. Yuming Huo for his invaluable advice and assistance with the neural network coding process.

Unveiling Hidden Musical Semantics: Compositionality in Music Ngram Embeddings

Posted on 25 June 2025 by kerry.turcsi

PGR JGI Seed Corn Funding Project Blog 2023/24: Zhijin Guo

Introduction

The overall aim of this project is to analyse music scores by machine learning. These of course are different from sound recordings of music, since they are symbolic representations of what musicians play. But with encoded versions of these scores (in which the graphical symbols used by musicians are rendered as categorical data) we have the chance to turn these instructions in various sequences of pitches, harmonies, rhythms, and so on.

What were the aims of the seed corn project?

CRIM concerns a special genre of works from sixteenth century Europe in which a composer took some pre-existing piece and adapted the various melodies and harmonies in it to create a new but related composition. More specifically, the CRIM Project is concerned with polyphonic music, in which several independent lines are combined in contrapuntal combinations. As in the case of any given style of music, the patterns that composers create follow certain rules: they write using stereotypical melodic and rhythmic patterns. And they combine these tunes (‘soggetti’, from the Italian word for ‘subject’ or ‘theme’) in stereotypical ways. So, we have the dimensions of melody (line), rhythm (time), and harmony (what we’d get if we slice through the music at each instant.

A network of musical notations — *Figure 1. An illustration of music graph, nodes are music ngrams and edges are different relations between them.* *Image generated by DALL·E*.

We might thus ask the following kinds of questions about music:

Starting from a given composition, what would be its nearest neighbour, based on any given set of patterns we might chose to represent? A machine would of course not know anything about the composer, genre, or borrowing involved in those pieces, but it would be revealing to compare what a machine might tell us about this such ‘neighbours’ in light of what a human might know about them.

What communities of pieces can we identify in a given corpus? That is, if we attempt to classify of groups works in some way based on shared features, what kinds of communities emerge? Are these communities related to Style? Genre? Composer? Borrowing?

In contrast, if we take the various kinds of soggetti (or other basic ‘words’) as our starting point, what can we learn about their context? What soggetti happen before and after them? At the same time as them? What soggetti are most closely related to them? And through this what can we say about the ways each kind of pattern is used?

Interval as Vectors (Music Ngrams)

How can we model these soggetti? Of course they are just sequences of pitches and durations. But since musicians move these melodies around, it will not work simply to look for strings of pitches (since as listeners we can recognize that G-A-B sounds exactly the same as C-D-E). What we need to instead is to model these as distances between notes. Musicians call these ‘intervals’ and you could think of them like musical vectors. They have direction (up/down) and they have some length (X steps along the scale).

Here is an example of how we can use our CRIM Intervals tools (a Python/Pandas library) to harvest this kind of information from XML encodings of our scores. There is more to it than this, but the basic points are clear: the distances in the score are translated into a series of distances in a table. Each column represents the motions in one voice. Each row represents successive time intervals in the piece (1.0 = one quarter note).

An ngram for a section of music — *Figure 2. An example of ngram: [-3, 3, 2, -2], interval as vectors.*

Link Prediction

We are interested in predicting unobserved or missing relations between pairs of ngrams in our musical graph. Given two ngrams (nodes in the graph), the goal is to ascertain the type and likelihood of a potential relationship (edge) between them, be it sequential, vertical, or based on thematic similarity.

Sequential is tuples that come near each other time. This is Large Language Model which computes ‘context’. LLM then produces the semantic information that is latent in the data.

Vertical is tuples that happen at the same time. It is ANOTHER kind of context.

Thematic is based on some measure of similarity.

Upon training, the model’s performance is evaluated on a held-out test set, providing metrics such as precision, recall, and F1-score for each type of relationship. The model achieved a prediction accuracy of 78%.

Beyond its predictive capabilities, the model also generates embeddings for each ngram. These embeddings, which are high-dimensional vectors encapsulating the essence of each ngram in the context of the entire graph, can serve as invaluable tools for further musical analysis.