Can boba
flavor be captured?

Abstract

Computational gastronomy has long represented food through shared flavour compounds and recipe co-occurrence, from the original flavour network to FlavorGraph. In earlier work we showed that such embeddings already encode interpretable culinary dimensions, spanning taste, texture, geography, processing and culture. Epicure took this a step further: three sibling models trained on 4.1 million recipes in seven languages, each weighting chemistry and cooking practice differently. You can ask either “what do I cook this with?” or “what can substitute this?” (with mixed results), and steer from the familiar to the unexplored with a continuous dial. Rotate chocolate towards sweet baking, and the chemistry-driven model lands on a neighbourhood of red bean, matcha and purple sweet potato. I’ll talk about how we constructed these models, what we are aiming to do in future work, and how tools like this might sit alongside personal and cultural interpretation.

Bio

Dr Jakub Radzikowski is Culinary Education Designer (Principal Teaching Fellow) in Chemistry at Imperial College London. Since 2019 he has co-developed Chemical Kitchen, which teaches laboratory skills through professional cooking. The approach has grown into Medical and Biomedical Kitchen courses and a partnership with TUM, and its evaluations show the least confident students gain the most. He also co-developed the pandemic “Lab in a Box”, part of an RSC prize-winning project that is now held by the Science Museum Group. A molecular biologist (PhD, Groningen) and Le Cordon Bleu-trained chef who cooked in Michelin-starred restaurants, he developed the Epicure ingredient embeddings and supervises doctoral research in computational gastronomy for sustainable protein. He is a Senior Fellow of the Higher Education Academy.