PowerGrow: Growth-Centric Systems

Updated 9 July 2026

PowerGrow is a multifunctional concept spanning plant-centric systems and power-grid synthesis, using measurable growth states to trigger constrained interventions.
In horticulture, it enables applications like image-based fertigation, energy-aware lighting control, and digital twins that optimize resources and crop outcomes.
In power systems, PowerGrow supports co-generative synthesis of grid topologies and nodal dynamics, achieving high power-flow convergence and operational feasibility.

PowerGrow is a designation used in the literature for several technically distinct systems concerned with growth-aware sensing, control, and synthesis. In plant-oriented work, the name is applied to image-based fertigation, electricity-aware lighting control, plasma-activated water treatment, magnetopriming, plant-level digital twins, and plant-powered robotics; in power systems, it denotes a co-generative framework for synthesizing grid structure and nodal dynamics (Ahmad et al., 2013, Lork et al., 2020, Bhattacharjee et al., 24 Oct 2025, Mayborne et al., 1 Jun 2026, Murakami et al., 2024, He et al., 29 Aug 2025). This suggests a recurring motif: measurable state variables—visual traits, physicochemical signatures, biomass estimates, or graph-temporal embeddings—are used to drive constrained intervention.

1. Scope of the term

A common misconception is that PowerGrow denotes a single platform. In the cited literature, it does not. The designation spans multiple, otherwise unrelated, research programs.

Context	Core technical aim	Representative source
Greenhouse fertigation	Use plant-image feedback to trigger pump actuation	(Ahmad et al., 2013)
Indoor lighting control	Minimize electricity cost while maintaining growth	(Lork et al., 2020, Abbaspour et al., 4 Oct 2025, Daniels et al., 2023)
Physicochemical plant treatment	Modify growth via PAW or magnetopriming	(Bhattacharjee et al., 24 Oct 2025, Mehrabifard et al., 15 Nov 2025, Ferroni et al., 2023)
Plant-level estimation	Infer biomass and forecast short-horizon growth	(Mayborne et al., 1 Jun 2026)
Biohybrid actuation	Use plant growth as a mechanical actuator	(Murakami et al., 2024)
Power-system generation	Jointly synthesize feasible grids and dynamic loads	(He et al., 29 Aug 2025)

Only one cited work uses PowerGrow as the paper title itself, namely the power-grid synthesis framework (He et al., 29 Aug 2025). The plant-oriented uses are implementation labels or synthesized design concepts attached to specific sensing, modeling, and control pipelines.

2. Image-based plant feedback and automatic fertigation

In greenhouse cultivation, PowerGrow appears as a direct descendant of the Speaking Plant Approach (SPA), in which irrigation or fertigation decisions are based on feedback from the plant rather than only on environmental or root-zone sensors. The tomato study in a sterilized greenhouse used CCD imaging to monitor plant height under three fertigation electrical conductivity regimes—1.0–1.5 mS, 2.5–5.0 mS, and 10.0–12.5 mS—and to detect canopy shrinkage associated with wilt (Ahmad et al., 2013).

The growth-monitoring component extracted height in pixels from still images captured every three days against a red background. Camera-to-plant distance was increased from 30 cm to 170 cm in 10 cm increments every three days, and the analysis considered a “distance factor” so that height remained comparable over time. Average height increased non-linearly in all groups; the over-fertilized group grew fastest, the normal-fertilized group next, and the under-fertilized group slowest. Individual image analysis became impractical after about 43–49 days because neighboring leaves overlapped.

The control component used a representative plant placed before a red cloth and monitored canopy width at fixed intervals. Wilt degree was defined from width shrinkage relative to a morning “fresh” reference,

$\text{wilt\_degree}(t)=\frac{w_0-w(t)}{w_0},$

with watering triggered only when width shrinkage reached at least 2% and the current width was smaller than the previous width, indicating ongoing shrinkage. When both conditions were met, the pump was turned ON for 3 minutes and then OFF. In the 14-day automatic phase, images were captured every 30 minutes from 8:00 to 17:00, and the image-based controller reduced average nutritive water use from 101.6 liters/day under timer control to 17.3 liters/day, a reduction of more than 80%, while maintaining plant height development comparable to timer control (Ahmad et al., 2013).

The method is technically narrow. It assumes one representative plant can govern irrigation for 60 plants, requires isolated plant images with a contrasting background, and uses canopy width shrinkage as a proxy for water deficiency only. The paper explicitly notes that other stress indicators, such as leaf color for micronutrient status, were suggested as future work rather than implemented.

3. Lighting optimization and crop-level optimal control

In indoor farming, PowerGrow is associated with energy-aware lighting control under explicit crop-growth models. One lettuce study implemented a sensing–modeling–optimization loop in an indoor, air-conditioned room held at 25°C, using three levels of NFT hydroponic beds, top-mounted hourly imaging, ISL29125 lux sensors calibrated to PPFD, DHT22 temperature and humidity sensors, and Raspberry Pi 3 edge control (Lork et al., 2020). Leaf area was estimated from top-view images by modified K-means clustering with homography-based rectification, so that

$L(t)=s\cdot N(t),$

where $s$ is the calibrated pixel-area factor and $N(t)$ is the number of segmented plant pixels.

Growth was then modeled by a two-hidden-layer feedforward neural network with ReLU activations, dropout $p=0.5$ , Adam optimization, and MSE loss for 2000 epochs. Inputs were min–max normalized PPFD $_{\text{red,avg}}$ , PPFD $_{\text{blue,avg}}$ , EC, pH, and time since transplanting. Relative to a baseline, the genetic algorithm that optimized hourly red and blue setpoints over a 15-day horizon achieved 40–52% reductions in electricity cost and, in the high-revenue setting, increased final leaf area by about 6% while still cutting cost by about 41% (Lork et al., 2020).

A later indoor-farming formulation replaced offline schedule search with model predictive control augmented by transformer-based 24-hour forecasts of electricity prices and solar radiation. For a one-hectare greenhouse with lettuce and Ontario market data, the controller optimized light intensity and photoperiod subject to DLI, PPFD, photoperiod, and dark-interval constraints derived from plant experiments. Reported annual outcomes were a cost reduction of $318,400 (20.9%), a peak load decrease of 1.6 MW (33.32%), and total energy savings of 1890 MWh (20.2%) against a baseline recipe (Abbaspour et al., 4 Oct 2025).

A related vertical-farm study formulated daily optimal control of temperature, drought stress, and radiation using a smoothed hybrid crop model adapted from SIMPLE. For wheat cultivar “Batten,” the annualized optimization converged to an optimal cropping period of 102 days, yielding annual biomass of about 11.02 kg m $^{-2}$ y $^{-1}$ versus 10.22 kg m $^{-2}$ y $L(t)=s\cdot N(t),$ 0 for a constant-setpoint baseline, while the per-cycle cost proxy decreased from 149.21 €/kg to 104.55 €/kg (Daniels et al., 2023). Taken together, these works span data-driven scheduling, constrained MPC, and stage-aware optimal control; they also share an important limitation: plant circadian dynamics, stress from aggressive duty cycles, and richer spectral effects are only partially modeled.

4. Physicochemical growth interventions

PowerGrow is also used for interventions that alter plant growth by modifying the chemical state of water or seeds before or during cultivation. In one Chrysanthemum study, a helium–air micro-plasma jet operated at He:air = 14:1, 10 kHz, and 14 kV peak-to-peak was used to generate plasma-activated water (PAW). The reported optimum condition was 12 mL of de-ionized water treated for 40 minutes, yielding nitrate of about 10 ppm, pH of about 5.6, the highest observed ORP, and increased EC/TDS. Over two weeks, saplings watered daily with this PAW reached 10.2 cm in height, compared with 8.0 cm for deionized water and 7.1 cm for tap water, while soil fertility at day 14 was 2580 $L(t)=s\cdot N(t),$ 1S/cm versus 900 and 795 $L(t)=s\cdot N(t),$ 2S/cm for the controls (Bhattacharjee et al., 24 Oct 2025).

A comparative lettuce study examined PAW generated by transient spark, fountain dielectric barrier discharge, and microwave plasma. The three systems produced markedly different chemistries: transient spark yielded the highest H $L(t)=s\cdot N(t),$ 3O $L(t)=s\cdot N(t),$ 4 and NO $L(t)=s\cdot N(t),$ 5 at near-neutral pH; microwave plasma yielded a nitrate-dominant PAW with pH $L(t)=s\cdot N(t),$ 6, ORP $L(t)=s\cdot N(t),$ 7 mV, EC $L(t)=s\cdot N(t),$ 8 $L(t)=s\cdot N(t),$ 9S/cm, and NO $s$ 0 of about 6.91 mM. In six-week soil growth, microwave PAW produced the largest biomass gains, with fresh weight multiplied by 1.31 and dry weight by 1.62 relative to tap water. The study reported Pearson correlations of $s$ 1 for NO $s$ 2 versus fresh weight and $s$ 3 for NO $s$ 4 versus dry weight, whereas transient spark and fountain DBD more strongly affected early shoot elongation and pigment composition (Mehrabifard et al., 15 Nov 2025).

Magnetopriming constitutes a distinct physical intervention. Ferroni et al. exposed maize seeds to static magnetic fields from 50 to 350 mT for 1 hour and found a clear optimum at 150 mT. On day 10, average plantule total length was $s$ 5 cm at 150 mT versus $s$ 6 cm in control, corresponding to a 108.9% increase; germinative energy and germinative power remained near 93–94% and 95–96%, respectively, with no significant differences from control (Ferroni et al., 2023). The mechanistic basis remains unsettled. The literature cited in the study discusses the Radical Pair Mechanism, biogenic magnetic nanoparticles, and liquid-crystal realignment, but none is established as definitive.

Across these interventions, the principal controversies concern mechanism and reproducibility rather than the reported directional effects. PAW studies often leave storage stability and full RONS kinetics unresolved, while magnetopriming studies remain sensitive to exposure window, seed lot, and environmental heterogeneity.

5. Measurement-driven digital twins

In hydroponic lettuce production, PowerGrow is used to denote a plant-level digital twin architecture that continuously updates a growth model using measurements. The cited implementation integrated NFT hydroponics, canopy-level sensing of temperature, CO $s$ 7, and PAR, and hourly overhead RGB-D imaging from an Intel RealSense D405 mounted on a FarmBot Genesis v1.7 gantry. Nutrient conditions were maintained at EC 1300–1700 $s$ 8S/cm and pH 5.8–6.0 (Mayborne et al., 1 Jun 2026).

The measurement model was a two-stream CNN receiving RGB and depth channels separately and regressing fresh mass. DenseNet121 backbones for both streams, followed by an MLP, gave the best performance; HSV color segmentation improved accuracy further. Using a dataset of about 1308 RGB-D images from 125 butterhead lettuce plants, the system reduced RMSE from 1.91 g to 1.56 g with color segmentation, and then to 1.45 g with DenseNet121 backbones, with $s$ 9. This is consistent with the reported ability to estimate mass within about 1.5 g of ground truth (Mayborne et al., 1 Jun 2026).

Forecasting used the NiCoLet B3 lettuce model, a grey-box formulation with two carbon compartments, environmental drivers $N(t)$ 0, and plant fresh mass as the measurable output. On held-out plants, calibrated NiCoLet forecasts maintained mean absolute errors of about 1.9–2.2 g across 1–4 day horizons, outperforming a simple exponential early-growth model at longer horizons and proving more stable than the tested LSTM sequence model. The resulting architecture is measurement-driven in a strict sense: current biomass is inferred non-destructively, uncertain model parameters are updated from measurements, and short-horizon yield trajectories are forecast from the fused state estimate (Mayborne et al., 1 Jun 2026).

The main limitations are dataset scale, domain shift across cultivars or sites, and canopy occlusion in denser plantings. The cited study therefore positions the architecture as practical for hydroponic lettuce under controlled geometry rather than as a universally calibrated crop twin.

6. Plant growth as actuation

A different use of the growth-centered PowerGrow concept treats plants not as the controlled object but as the actuator itself. In the plant-robotics study, radish sprouts were characterized as light-driven, self-powered actuators capable of producing both displacement and force. After 40 hours, average lengths reached 46.4 mm in dark conditions and 16.3 mm under illumination; peak sprouting forces were $N(t)$ 1 mN in the dark and $N(t)$ 2 mN under illumination (Murakami et al., 2024).

These measurements supported two biohybrid robots. The first was a rotational mobile robot in which four radish sprouts, arranged at 45° intervals in a cylindrical frame, generated tangential ground contact sufficient to overcome rolling resistance of about 5 mN. It achieved a horizontal displacement of 14.6 mm in the dark and 10.4 mm under illumination, with average speed until 15 hours of 0.8 mm/h and 0.7 mm/h, respectively. The second was a gripper using phototropism: inner LEDs induced inward bending for grasping, outer LEDs induced outward bending for release, and the device picked and placed a 0.1-g sponge (Murakami et al., 2024).

The work is explicit about its operating envelope. Reported power densities were $N(t)$ 3 W/kg in the dark and $N(t)$ 4 W/kg under illumination on a seed-mass basis, with hour-scale response times. This rules out comparison with conventional fast actuators except in a narrowly ecological or biohybrid sense. The significance lies instead in model-based exploitation of slow biological growth for locomotion and manipulation.

7. PowerGrow as a power-grid synthesis framework

In power systems, PowerGrow refers to a co-generative framework for synthesizing grid topology, branch attributes, bus properties, and nodal dynamics while preserving operational realism. The method factorizes the joint distribution of structure and dynamics into a hierarchy,

$N(t)$ 5

where $N(t)$ 6 is adjacency, $N(t)$ 7 bus features, $N(t)$ 8 edge attributes, and $N(t)$ 9 latent load embeddings learned by an LSTM autoencoder (He et al., 29 Aug 2025).

Structural synthesis is performed by hierarchical graph beta-diffusion in bounded $p=0.5$ 0 spaces, while time-series loads are encoded into a latent dimension $p=0.5$ 1 before diffusion and decoding. The framework was evaluated on IEEE 14-bus and European 36-bus benchmarks, using 878 feasible 14-bus variants and 813 feasible 36-bus variants after AC power-flow screening. On the 36-bus system, reported sample-generation time for a complete topology-plus-load instance was 0.525 s, compared with 0.635 s for EDP-GNN, 0.765 s for GruM, 2.620 s for GDSS, and 46.255 s for an optimization-based post-process (He et al., 29 Aug 2025).

Operational validity is assessed post hoc rather than enforced explicitly during sampling. Even so, the framework achieved a 98.9% AC power-flow convergence rate, an average feasibility score of 0.967, and 73.19% N-1 contingency resilience, compared with 72.91% for the reference system and 65.24% for random-walk variants. The central technical claim is not that physical constraints are hard-coded into the denoising loop, but that dependence decomposition, bounded beta-diffusion, and training on feasible data materially improve the probability of generating power-flow-convergent and operationally credible cases (He et al., 29 Aug 2025).

Its limitations are correspondingly clear. Feasibility is emergent rather than guaranteed; the hierarchical factorization imposes conditional independence assumptions that may omit residual dependencies; and empirical validation is confined to 14-bus and 36-bus settings. Within that scope, however, PowerGrow is a concrete example of graph-temporal co-generation rather than a generic label for grid anonymization.

In aggregate, PowerGrow is best understood not as a single method but as a recurring label for systems that couple measured growth-related states to structured intervention. In controlled horticulture, this coupling appears as visual feedback, crop-growth models, plasma chemistry, or magnetic pre-treatment; in biohybrid robotics, it appears as direct mechanical actuation by living tissue; in power systems, it appears as hierarchical co-generation of network structure and dynamics. The term therefore denotes a family of growth-centered technical programs rather than a unified theory.