Plant Calcium Signals and VAE
This work was carried out as my Bachelor's Thesis within a research group at CNR (National Research Council of Italy), in collaboration with the RFX Consortium (Cortese et al., 2021; Antoni, Cortese, and Navazio, 2025). The central question was how to non-invasively control plant behaviour — for example triggering an early immune response or inducing deeper root growth ahead of a drought — without any genetic modification.
Biological Background
Most plant responses to external stimuli are encoded through calcium (Ca²⁺) signalling cascades. Replicating a specific Ca²⁺ signal would effectively induce the corresponding plant response, even in the absence of the original triggering event. This opens a non-invasive avenue for plant control.
It is known that plasma-activated water (PAW) — water exposed to a non-thermal plasma discharge — induces a characteristic Ca²⁺ transduction pattern in plants upon watering. Crucially, the shape of this signal depends on PAW production parameters such as exposure time and plasma power. The challenge is therefore an inverse problem: given a desired target Ca²⁺ signal, find the PAW parameters that produce it.
Approach
To approximate this complex and noisy mapping, we trained a beta-VAE on a dataset of recorded Ca²⁺ signals paired with their corresponding PAW parameters. The regularised variational objective encourages the latent space to disentangle the main factors of variation in the transduced signal, making interpolation and nearest-neighbour search geometrically meaningful.
At inference time, we encode the target Ca²⁺ signal into the same latent space and apply k-NN to retrieve the closest training examples. The PAW parameters associated with those neighbours are then used as the recommended production settings — the PAW generated under those conditions, when applied to the plant, should induce a signal closely resembling the target.
