Plants, like animals, carry information about their physiological state through electrical signals that can be recorded with electrodes inserted into conducting tissue. Earlier machine-learning models could predict specific stress states from these signals, but each was built for a single task and did not transfer across species, environments, or conditions. That limitation made it hard to scale plant monitoring to the diversity found in agriculture, forestry, and horticulture.
Bombadil, a foundation model described in a new preprint, is pretrained without labels on more than 4 million hours of electrical signals from 100 plant species across ten environments, including laboratories, greenhouses, orchards, and forests. By learning from such a broad dataset, the model captures fundamental features of plant electrical activity rather than task-specific patterns.
From those same signals, Bombadil can track changes in seven physiological indicators of plant stress, even for species and stressors that had no labels in its training data. This suggests the model generalizes beyond its pretraining set, potentially enabling stress detection in plants it has never encountered. The authors report the work in a 25-page paper with 13 figures and 11 tables, submitted to arXiv on 19 October 2026.