How the fruit fly connectome can help autonomous materials discovery
- 1Department of Materials Science and Engineering
- 2University of Utah
Abstract
Autonomous laboratories for two-dimensional semiconductors pair a recipe planner with instruments that return in situ spectra every few seconds. Some decisions are too fast for the planner: stop a failing selenization anneal, flag a flake that is not the logged material, or end an ion irradiation at a set vacancy density, from a few examples, on a small computer at the instrument. The fruit fly takes decisions of this form with two structures whose wiring is measured at synapse level: the mushroom body, a one-shot memory for odours, and the ventral nerve cord, which turns a few descending commands into local reflexes. We asked which instrument jobs models wired from these connectomes do better than engineered tools. A mushroom-body model of 685 projection neurons and 5,177 Kenyon cells read one Raman band per glomerulus; over 20 paired seeds, on simulated selenization Raman streams and on Ar + irradiation anchored in published in situ spectra, it lost the fewest flakes when one failed run taught a stop rule (3.3 of 30, against 3.9 for a nearest neighbour), read vacancy density from its sparse code with less error than from the raw spectrum (5.5 % against 6.4 %), and ran in 0.30 ms on a Raspberry Pi 5 within 152 kB. A stop decided at the furnace, as the nerve cord decides reflexes, prevented the losses of a planner that needed 60 s per turn (0.15 against 10.7 flakes of 30). It did not choose recipes. In three animals and after training, measured wiring matched a degree-preserving shuffle, and the nerve-cord wiring matched a ridge regression: the connectome contributes an architecture and wiring statistics, not a partner map. As hardware, the fixed wiring can be binary and each plastic synapse needs about 64 levels. The connectome supplies a low-cost reflex layer under the planner, not a replacement.
Related
- Autonomous in-silico inorganic materials discovery via multi-agent physics-aware scientific reasoning
- Hierarchical Multi-agent Large Language Model Reasoning for Autonomous Heterogeneous Catalyst Discovery
- Accelerating the discovery of nanoporous gas separation materials via Coulomb-enhanced classical density functional theory
- Scalable Thermoset Informatics Integrated with Autonomous Experimentation for On-Demand Discovery
- Evaluating transfer across modalities in autonomous materials systems
BibTeX
@misc{ackolf2026fruit,
title = {{How the fruit fly connectome can help autonomous materials discovery}},
author = {Kinston Ackölf and Taylor D. Sparks},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010667/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010667/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - How the fruit fly connectome can help autonomous materials discovery
AU - Kinston Ackölf
AU - Taylor D. Sparks
PY - 2026
DA - 2026/10/19
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010667/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010667/v1
AB - Autonomous laboratories for two-dimensional semiconductors pair a recipe planner with instruments that return in situ spectra every few seconds. Some decisions are too fast for the planner: stop a failing selenization anneal, flag a flake that is not the logged material, or end an ion irradiation at a set vacancy density, from a few examples, on a small computer at the instrument. The fruit fly takes decisions of this form with two structures whose wiring is measured at synapse level: the mushroom body, a one-shot memory for odours, and the ventral nerve cord, which turns a few descending commands into local reflexes. We asked which instrument jobs models wired from these connectomes do better than engineered tools. A mushroom-body model of 685 projection neurons and 5,177 Kenyon cells read one Raman band per glomerulus; over 20 paired seeds, on simulated selenization Raman streams and on Ar + irradiation anchored in published in situ spectra, it lost the fewest flakes when one failed run taught a stop rule (3.3 of 30, against 3.9 for a nearest neighbour), read vacancy density from its sparse code with less error than from the raw spectrum (5.5 % against 6.4 %), and ran in 0.30 ms on a Raspberry Pi 5 within 152 kB. A stop decided at the furnace, as the nerve cord decides reflexes, prevented the losses of a planner that needed 60 s per turn (0.15 against 10.7 flakes of 30). It did not choose recipes. In three animals and after training, measured wiring matched a degree-preserving shuffle, and the nerve-cord wiring matched a ridge regression: the connectome contributes an architecture and wiring statistics, not a partner map. As hardware, the fixed wiring can be binary and each plastic synapse needs about 64 levels. The connectome supplies a low-cost reflex layer under the planner, not a replacement.
ER -