Translucent AI

Translucent lets AI agents do experimental engineering, from understanding lab data to deciding what experiment to run next. We built it using Meta's Muse Wonder. Give it raw test records and a goal, and it figures out what was measured, writes the analysis, tests physical models, and runs experiments. When a model fails to explain the data, Muse decides what to test next, runs a simulated experiment, and uses the result to revise its model. We demonstrated this on battery fast-charging research. Muse found that two cells behaved differently, chose new tests, and used simulated lab results to find a feasible charging protocol. In a separate benchmark, it found designs 5 to 12 simulated charging minutes faster than Bayesian optimization with the same eight-experiment budget. Translucent also uses Muse Spark and SAM 3.1 to measure changes in real laboratory microscopy footage. We found real improvements as shown in our research paper and code. The goal is to let AI do more of the scientific work between an observation and the next experiment. Physical battery validation is still outstanding.