Plants, like animals, use electrical signals to carry information about their physiological state across their entire body, from roots to branches. These signals have enabled devices in recent decades capable of "listening" to a plant and using its impulses to move a pot toward light or trigger an alarm when soil moisture drops below a threshold. What changed radically in October 2026 is the ability to interpret these signals at scale. The Electric Plant Company, a San Francisco startup that emerged from stealth this week, unveiled Bombadil — the first foundation model of machine learning dedicated to plant electrophysiology — trained on over 4 million hours of electrical signals recorded from 100 different species across ten distinct environments, including laboratories, commercial greenhouses, orchards, and natural forests.
The model was developed by a team led by CEO Justin Mast, founder of the earlier online plant company Bloomscape, and chief scientist Lee von Kraus, a neuroscientist who worked with DARPA and the U.S. Environmental Protection Agency. Their approach is analogous to what transformed natural language processing and computer vision: instead of training task-specific models for each application — as many prior research groups had done, each focused on predicting only drought, or only a single species — they built a generalist backbone that learns the fundamental structure of plant electrical signals in a self-supervised manner, without requiring labeled data.
Electric Plant's hardware consists of a sensor called "Bud," a low-cost device that attaches to a plant's stem and continuously records raw electrical signals. The electrodes are inserted deep enough to penetrate xylem and phloem — the channels that conduct electrical signals in plants. Wire gauge varies by species: for fleshy-stemmed plants, 28-gauge wire with Ag/AgCl coating or stainless steel is used; for woody plants, stainless steel screws are employed. The sensor transmits raw data to the cloud, where Bombadil decodes the signals and estimates seven physiological stress markers: photosystem II efficiency, leaf temperature relative to ambient air, Crop Water Stress Index (CWSI), among other parameters established in plant physiology literature.
What makes Bombadil genuinely different from prior plant electrophysiology models — as detailed in the technical paper published on arXiv (arXiv:2610.26597) by James Kirk, Piotr Bednarski, Filip Twardy, Justin Mast, and Lee von Kraus — is its generalization capability. The model was able to estimate physiological indicators in species it had never seen during pretraining — a result the authors say substantially outperforms any point model trained exclusively for a single species or task. This is particularly relevant to the global agricultural landscape, where hundreds of thousands of species are cultivated, each with distinct physiological responses to different stressors.
Electric Plant raised $11.2 million in a seed round led by Onto Ventures and Climactic, with participation from Ponderosa Partners, Endeavor Catalyst, and other angel investors. The Bombadil model is capable of interpreting electrical signals in approximately 125 different species — from ferns and redwoods to tomatoes and dahlias — and the company plans to continuously expand its dataset and capabilities. Founder Justin Mast told TechCrunch that "we know the model will surpass a lot of that [traditional instrumentation]".
The science behind Bombadil is not new. Plant electrophysiology has decades of published research supporting the idea that plants produce electrical signals in response to environmental stimuli. What is new is the scale and the deep learning approach applied to the problem. Until recently, each machine learning model was trained to predict a single stress state in a single species, in a single environment. Bombadil shifts this paradigm by learning the fundamentals of plant electrical signals en masse, in a generalist fashion, and then applying that knowledge to new species and new tasks through additional "heads" that connect to the pretrained backbone.
The authors also highlight important limitations. None of the recordings in the dataset last longer than three months per plant, meaning seasonal variation and the full life cycle of plants are not fully represented. Future extensions of the model include collecting data across more seasons, adding new physiological markers, and evolving the architecture to read multiple sensors simultaneously rather than blending estimates post hoc. But the milestone already exists: for the first time, a deep learning model can generalize across radically different plant species to infer their physiological state from pure electrical signals.
What does plant electrophysiology with AI reveal about the future of agriculture? Perhaps more important than the immediate practical application is the shift in perspective it represents: plants are not passive machines that merely "absorb" light and water — they are active electrophysiological systems that continuously communicate their state to their environment, and now we finally have the technical capacity to "listen" in a systematic and scalable way. The question is whether, just as language models learned to produce coherent text from sequences of symbols, plant electrophysiology models will learn to produce accurate predictions about plant behavior and needs — and whether, at some point in the future, it will become possible to literally talk to them.
Sources: arXiv, TechCrunch, BusinessWire
✓ Independent sources cross-checked and verified before publishing
