Research at EPCC: weather as music

Takeshi Matsumura, a Data Science, Technology and Innovation MSc student, and I presented our paper on AI-driven weather sonification at EvoMUSART 2026, one of four conferences gathered under the Evostar umbrella.

Weather sonification via a latent emotion space

Conventional weather apps are visual by design, which creates a genuine barrier for blind and visually impaired (BVI) people and many neurodivergent users. Our project represents the first steps towards conveying the weather forecast through music alone; it is designed to be intuitive, without any need for the listener to use a screen or learn what the mappings mean. It also speaks to broader ambitions: that music may reveal patterns, trends, or anomalies in data that conventional statistical or visual methods miss, given the human auditory system is finely attuned to temporal and tonal variation. Thus, data sonification may prove a valuable analytical tool for any researcher and, through its accessibility for BVI and neurodivergent users, widen participation in STEAM subjects (science, technology, engineering, arts and mathematics). 

The central challenge is mapping meteorological data to musical parameters in a way that feels natural rather than arbitrary. Our ML system does this by routing data through an intermediate psychological space, grounded in Russell's circumplex model of affect, which organises emotional states along valence (pleasant to unpleasant) and arousal (calm to excited) axes. 

Weather conditions from the Met Office's high-resolution 2km numerical weather prediction model are mapped to coordinates in this space, which then drive musical output: tempo, note duration, pitch range, and mode. A Variational Autoencoder learns the latent structure of the emotion space, while a feed-forward neural network handles the mapping to musical features. 

Listening evaluations showed participants could broadly distinguish good weather from bad weather, though finer distinctions proved harder, attributed in part to Edinburgh's moderate climate not producing sufficiently contrasting emotional mappings. Our longer-term goal is a deployable app: press play and know whether to reach for your brolly.

Get involved: Weather music listening evaluation

Our listening evaluation is still open and you are warmly invited to take part in our study 'Weather Sonification Evaluation'. The study takes only a few minutes: you will hear short pieces of music generated from real weather forecast data and be asked for your impressions. No musical training is required. Participation is voluntary, responses can be anonymous, and data will be used for research purposes only. 

Evostar 2026

Around 180 attendees gathered in Toulouse earlier this year for Evostar, which brings together four co-located conferences on evolutionary and nature-inspired computation: EuroGP (Genetic Programming), EvoApplications (applications of evolutionary computation), EvoCOP (combinatorial optimisation), and EvoMUSART (AI in music, sound, art and design, now in its 15th year).

Our EvoMUSART conference experience

Our session opened the EvoMUSART track and, despite a slightly late start and no microphones, the audience was visibly engaged throughout, with many taking photographs of the slides. The poster session later drew a steady queue, which was a welcome surprise.

Other EvoMUSART talks covered audio patch search, quantum-inspired latent spaces for music generation, and the broader cultural role of generative AI, framed not just as a production tool but as a socio-technical force reshaping how music is selected and valued. A talk on a project with a Malian musician exploring music and cultural identity produced a striking moment: when the researcher tried to write about the River Niger, or the neighbouring country of Niger, the AI refused on grounds of racist language. A reminder that AI bias is not always subtle.

A highlight of the final day was discovering that a speaker on astronomical sonification was the author of the ML model that formed the basis of our own work. However, their approach prioritises scientific fidelity over intuitive listening, requiring trained audiences, while ours does the opposite.

The two Evostar keynotes and one invited talk were equally stimulating. Simon Lucas (Queen Mary University of London) demonstrated that LLMs lose 100% of deterministic games like Connect Four or Monte Carlo Tree Search. They plan well but suffer from frequency bias and a "knowing-doing gap", making a strong case that simulation-based AI and evolutionary search remain indispensable. 

Jean-Baptiste Mouret, recipient of the 2025 Julian Francis Miller Award, argued that fitness-only optimisation is insufficient for open-ended search, with quality-diversity algorithms (occupying distinct niches rather than competing for a single optimum) offering a more creative path forward. 

Guy Theraulaz (CNRS, Toulouse) closed with a thrilling keynote on collective behaviour in fish schools, drawing entirely on physics and biology. A highlight was his team's elegant experiment: a virtual fish projected onto the side of a real fish tank using forced perspective, with the actual fish's eyes tracked to ensure the illusion felt real. 

Links

Evostar 2027 will be held in Mainz, Germany.

Matsumura, T., Pringle, G.J. (2026). Weather Sonification via a Latent Emotion Space: A Deep Learning Approach. In: Machado, P., Romero, J.J., Rebelo, S.M. (eds) Artificial Intelligence in Music, Sound, Art and Design. EvoMUSART 2026. Lecture Notes in Computer Science, vol 16523. Springer, Cham. https://doi.org/10.1007/978-3-032-24350-8_22