The experiment was conducted at the University of Georgia (College of Agricultural and Environmental Sciences, Department of Horticulture, Controlled Environment Agriculture Crop Physiology and Production Laboratory) located in Athens, Georgia, USA (33 93’11.36” N, 83 36’39.28” W). Lettuce (Lactuca sativa ‘Cherokee’) seeds (Johnny’s Selected Seeds, Waterville, ME, USA) were sown weekly into 96-cell trays containing a peat-perlite substrate (Fafard 1P; SunGro Horticulture, Agawam, MA, USA). The seedlings were then grown in a walk-in growth chamber under white LED light bars (RAY series with Physiospec indoor spectrum; Fluence Bioengineering, Austin, TX, USA) with a spectral distribution of 39% red, 40% green, 18% blue, and 3% far-red radiation. Under a 16-h photoperiod, canopy-level PPFD was maintained at 320 μmol·m^-2·s^-1, corresponding to a DLI of 18.4 mol·m^-2·d^-1. Plants were irrigated daily using an automated ebb-and-flow subirrigation system supplying a 15N-2.2P-12.4K fertilizer (Jack’s Professional^® LX 15-5–15 Cal-Mg LX; JR Peters, Allentown, PA, USA) at 100 mg·L^-1 N (nitrogen). During seedling cultivation, the growth chamber environment averaged 24.2 ± 0.3 °C for air temperature (Temp), 1.24 ± 0.06 kPa for vapor pressure deficit (VPD), and 831.6 ± 12.5 μmol·mol^-1 for carbon dioxide concentration (CO₂) (mean ± standard deviation [SD]).

After three weeks, seedlings were transplanted into 10-cm square containers using the same soilless substrate and moved to a glass-covered greenhouse in Athens, Georgia, USA (33°57′26.676″ N, 83°22′36.48″ W). Following one week of acclimation period in the greenhouse, the 4-week-old plants were used for two types of experiment: 1) data collection for the development of ML models predicting Φ_PSII, and 2) a validation experiment comparing biofeedback systems based on ML-predicted and directly measured CF values, with an additional constant PPFD control treatment. To ensure consistent plant age across experiments, this entire process for preparing plant materials was repeated weekly. Each week, a new batch of plants was used for the data collection or the validation experiment. During both the model development and validation experiments, all greenhouse plants were irrigated daily using the same subirrigation system with a fertilizer concentration of 200 mg·L^-1 N.

Over a period of six consecutive weeks, four lettuce plants were placed in an ebb-and-flow subirrigation tray (125 × 125 × 10 cm, W × D × H) and used per week for data collection. A multi-environment sensor (SM-600; Apogee Instruments, Logan, UT, USA) was installed at the center of the tray to monitor Temp, VPD, and CO₂ every 30 seconds, and 15-min averaged values were recorded. To regulate incoming sunlight intensity and maintain a DLI level appropriate for lettuce (≈ 16 mol·m^-2·d^-1), a 50% shade curtain was installed above the tray. Light levels were measured every 5 seconds using two pairs of quantum sensors: photosynthetically active radiation (PAR, 400–700 nm) and extended PAR (ePAR, 400–750 nm) sensors (SQ-500-SS and SQ-610-SS; Apogee Instruments, Logan, UT, USA). One set of each quantum sensor type was positioned directly next to the uppermost fully expanded leaf of each plant. Instantaneous light intensities measured at 15-minute intervals were designated PPFDi and ePPFDi, whereas PPFD15 and ePPFD15 denoted the average values over the preceding 15 minutes. Sub-daily accumulated light integrals, DLI_cum and eDLI_cum, were calculated at the same intervals to represent cumulative light received from dawn to the time of measurement.

CF was measured every 15 minutes using a multi-site pulse-amplitude modulated (PAM) fluorometer (MONITORING-PAM; Heinz Walz, Effeltrich, Germany). The uppermost fully expanded leaf of each plant was measured at the middle portion of the leaf lamina between secondary veins using four measuring heads, and data acquisition was automated using proprietary software (WinControl-3; Heinz Walz, Effeltrich, Germany).

All environmental sensors were connected to a datalogger (CR1000X; Campbell Scientific, Logan, UT, USA) for continuous data collection. Daily average environmental conditions and DLI during data collection and validation experiment are presented in Figure 1. All environmental variables and CF data were recorded simultaneously at 15-minute intervals (:00,:15,:30, and:45) throughout the 13.5 h photoperiod (06:45–20:15). Environmental variables, measured at the center of the tray, were shared among the four plants at each time point. A total of 5,079 data points from 24 lettuce plants were used to develop the model.

A supervised ML approach using multiple linear regression (MLR) was developed to predict Φ_PSII based on environmental and temporal variables. Time-of-day variable was included as a binary categorical variable with two levels, “beforenoon” and “afternoon”, created by splitting the data at 13:30, corresponding to the approximate solar midday at the experimental site during the study period. Initial variable screening was performed by examining the bivariate relationships between Φ_PSII and each continuous predictor (Table 1). Stepwise regression based on Akaike Information Criterion (AIC) was performed using the stepAIC function in the MASS package in R (v 4.5.0; R Foundation for Statistical Computing, Vienna, Austria). Three selection approaches, forward selection, backward elimination, and bidirectional stepwise selection, consistently indicated the exclusion of PPFDi. To further address multicollinearity, models were manually refined, and the final variable set was selected based on AIC and variance inflation factor (VIF < 5).

The final MLR model included the following predictors: ePPFDi, ePPFD15, Temp, VPD, CO₂, and Time-of-day. Variables were not normalized to preserve original units for direct use of regression coefficients in the validation experiment. The dataset was randomly split into training (70%) and testing (30%) sets. Five-fold cross-validation was conducted on the training data using the tidymodels package in R to assess model performance and stability. Hyperparameter tuning was not performed, and model performance was summarized by averaging evaluation metrics across folds. The model performance was evaluated using the root mean square error (RMSE), coefficient of determination (R²), and mean absolute error (MAE) on both the training and test datasets. Regression coefficients, along with their standard errors and P-values, are presented in Table 2. Variable importance was assessed using the vip package in R to identify the relative contribution of each predictor variable in the final model (Figure 2). Linearity and homoscedasticity were assessed using residuals-versus-fitted plots, while residual normality was evaluated using normal Q-Q plots. Multicollinearity in the final model was evaluated using VIF, and temporal autocorrelation was assessed using autocorrelation function (ACF) plots and the Durbin-Watson test.

A validation experiment was conducted in the same greenhouse to evaluate the applicability of the ML model for real-time supplemental lighting control. The experiment was conducted on a 9 m × 1.5 m bench divided into five blocks. Each block contained five experimental units (0.36 m × 1.5 m) separated by vertical aluminum panels to prevent light interference between treatments. Supplemental light treatments were randomized within each block. A 60% shade net was installed above the bench to achieve the DLI within an appropriate range by limiting excess solar radiation. Environmental conditions were measured using the same multi-environment sensor (SM-600; Apogee Instruments, Logan, UT, USA) positioned above the third block, between the supplemental LED lights and the shade net. Plants were irrigated daily using an automated ebb-and-flow system with the same nutrient solution as during the data-collection period.

White LED light bars (RAY series with Physiospec Greenhouse spectrum; Fluence Bioengineering, Austin, TX, USA) were installed above each experimental unit. Each bar was 1.1 m in length and had a spectral composition of 37% red, 39% green, 21% blue, and 3% far-red. Five quantum sensors (SQ-500-SS; Apogee Instruments, Logan, UT, USA) were positioned 25 cm below the LED bars, beside the plant canopy, and at the same height as the CF measurement spots in each treatment of the third block. These sensors measured the combined PPFD from sunlight and supplemental LED lighting every 15 seconds. All quantum and environmental sensors were connected to one of two dataloggers (CR1000X; Campbell Scientific, Logan, UT, USA) for Φ_PSII prediction and lighting control.

CF parameters were measured using two types of the PAM fluorometers (Heinz Walz, Effeltrich, Germany) for the validation experiment. Two MINI-PAM units were connected to the datalogger via RS-232 interface for real-time measurement of CF parameters used in the sensor-based lighting control. All measurements were conducted every 15 minutes during the 13-hour photoperiod and were automated using proprietary software either LoggerNet (Campbell Scientific) or WinControl-3 (Heinz Walz). Measurements were taken on the uppermost fully expanded leaves of each plant at the middle portion of the leaf, as described above. The CF parameters included Φ_PSII and ETR, which were calculated using the following equations (Hendrickson et al., 2004):

where F_s is the steady-state fluorescence yield under actinic light, Fm′ is the maximum fluorescence of a light-adapted leaf, and F_m is the maximum fluorescence of a dark-adapted leaf.

Two sensor-based biofeedback logic modules were implemented in the datalogger to control supplemental light intensity based on either ETR or Φ_PSII. Real-time CF measurements were obtained every 15 minutes during the photoperiod, and the datalogger computed the new target PPFD (PPFD_t) accordingly. For the ETR-based control, PPFD_t was calculated using the ratio between the target ETR (ETR_t) and current ETR (ETR_c) values as follows:

This approach assumes a linear relationship between ETR and PPFD under sub-saturating light levels. The conditional clause minimized unnecessary light adjustments when ETR_c was within ±5% of the target.

For the Φ_PSII-based control, PPFD_t was computed from the deviation between the target (Φ_{PSII, t}) and current Φ_PSII (Φ_{PSII, c}) values, using the slope (k) derived from the empirical relationship between PPFD and Φ_PSII based on the dataset collected for ML model development, as follows:

The slope k was calculated from the linear regression of Φ_PSII against PPFD (Φ_PSII = 0.7347 − 0.0002827 × PPFD), giving k = 0.0002827. This allowed conversion of a difference in Φ_PSII to an equivalent PPFD adjustment. Likewise, ± 1% threshold around Φ_{PSII, t} was applied to avoid over-reactive adjustments in PPFD.

The MLR model developed was implemented to predict the Φ_PSII in real time using environmental and temporal inputs. The model equation was applied as:

where Time-of-day is a binary variable (1 for before 13:30, 0 otherwise).

All input variables were collected in real time. Temp, VPD, and CO₂ were measured every minute, whereas ePPFD were measured every 15 seconds using ePAR sensors (SQ-610-SS; Apogee Instruments, Logan, UT, USA), installed only in the ML-based lighting treatments. ePPFD₁₅ was calculated as the average ePPFDi over the preceding 15 min, based on measurements taken every 15 s. The predicted Φ_PSII and ETR were then used to calculate PPFDt, following the same control logic as for sensor-based lighting control, with all computations performed in real time by the datalogger.

Each datalogger was connected to an analog output module (SDM-AO4A, Campbell Scientific, Logan, UT, USA) to control the LED light levels. The module sent 0–10 V direct current dimming signals to the LED drivers, each of which powered five LED bars from the five blocks. Supplemental lighting was provided for 13 h per day (07:00–20:00). To dynamically adjust light intensity in response to changing sunlight, the dimming signal was updated every 15 seconds based on the following equation:

where PPFD_t is the target PPFD calculated from the biofeedback system or constant PPFD treatment, and PPFD_c is the current light intensity measured by the quantum sensors.

At the beginning of the photoperiod, all treatments were adjusted to a total PPFD of 350 μmol·m^-2·s^-1 for 30 minutes to minimize transient fluctuations before stable operation of the biofeedback light control, as previously demonstrated (Nam et al., 2025). In sensor-based control, PPFD_t was adjusted every 15 min because saturating pulses during CF measurements can induce photoinhibition if applied too frequently. In contrast, the ML-based control predicted CF parameters using non-invasive environmental measurements, allowing PPFD_t to be adjusted every 15 seconds. Regardless of the PPFD_t adjustment interval, the dimming signal was recalculated every 15 seconds using PPFD_c to respond promptly to fluctuations in sunlight. In the constant PPFD treatment, PPFD_t was fixed at 350 μmol·m^-2·s^-1, and the dimming signal was directly adjusted every 15 s based on the ration of PPFD_t to PPFD_c.

All plants were harvested 15 days after the supplemental lighting treatments. Shoot fresh weight was recorded, and shoot dry weight was determined after drying at 80 °C for 72 h. Leaf chlorophyll and anthocyanin contents were measured on three uppermost fully expanded leaves per plant using handheld meters (CCM-200 plus and ACM-200 plus; Opti-Sciences, Hudson, NH, USA), and the mean value was used for each experimental unit.

A linear relationship between dimming signals and power consumption was determined using a power meter (P4400; P3 International Corporation, New York, NY, USA). Power consumption was measured at dimming signals ranging from 0 to 10 V in 1 V increments, with five replicates at each level. The resulting equation, Power consumption (W) = 7.93 × dimming signal (V) (R² = 0.99), was used to estimate instantaneous power use for the supplemental LED lighting, which was then integrated over time to calculate total electricity consumption per plant. Energy use efficiency (g·kWh^-1) was calculated as the ratio of shoot fresh and dry weight to the amount of electricity consumed.

Five supplemental lighting treatments were arranged in a randomized complete block design (RCBD) with five blocks (n = 5). The treatments consisted of two types of biofeedback regulation: a sensor-based control using real-time CF measurements and an ML-based predictive control. Each control strategy was implemented using either an ETR-based logic or a Φ_PSII-based logic. Accordingly, the treatments included sensor-based biofeedback control targeting an ETR of 100 μmol·m^-2·s^-1 (BFB-ETR 100) and ML-based control targeting the same ETR (ML-ETR 100), as well as sensor-based control maintaining a target Φ_PSII of 0.68 (BFB-Φ_PSII 0.68) and ML-based control targeting the same Φ_PSII value (ML-Φ_PSII 0.68). In addition, a constant PPFD treatment was included, where supplemental lighting was adjusted in response to changing sunlight to maintain a target PPFD of 350 μmol·m^-2·s^-1 (PPFD 350), representing a conventional light control strategy for comparison. Specific target ETR, Φ_PSII, and PPFD values were selected prior to the experiment, based on preliminary tests, to provide DLI levels sufficient for lettuce growth and achieve comparable DLIs across treatments.

All statistical analyses were performed using R (v4.5.0; R Foundation for Statistical Computing, Vienna, Austria). Linear mixed-effects models were applied, treating treatment as a fixed effect and block as a random effect. Harvest and energy use parameters were analyzed using one-way analysis of variance (ANOVA), and pairwise comparisons were performed with Tukey’s Honestly Significant Difference (HSD) test at a 95% confidence level.

Environmental conditions varied widely during the data collection period, providing a broad range for training the ML model (Figure 1). Temperature and CO₂ showed relatively modest variability, averaging 22.4 ± 2.5 °C (range 14.8–33.2 °C) and 483 ± 38 μmol·mol^-1 (range 325–605 μmol·mol^-1), respectively. In contrast, VPD and DLI exhibited substantial day-to-day fluctuations. VPD averaged 0.51 ± 0.43 kPa (range 0–3.1 kPa), and DLI averaged 15.9 ± 5.7 mol·m^-2·d^-1 (range 2.9–22.8 mol·m^-2·d^-1). Environmental conditions during the validation experiment overlapped with those observed during the data collection period but exhibited differences in their distribution and ranges (Supplementary Figure 1). Temperature averaged 25.2 ± 3.0 °C (range 20.0–33.1 °C), and VPD averaged 0.65 ± 0.56 kPa (range 0–2.5 kPa) during the validation experiment (Figure 1). CO₂ was moderately higher, averaging 549 ± 51 μmol·mol^-1 (range 380–680 μmol·mol^-1), but still overlapped with most values from the data collection period. DLI averaged 9.2 ± 3.3 mol·m^-2·d^-1 (range 1.7–12.6 mol·m^-2·d^-1) without supplemental lighting and 15.6 ± 1.8 mol·m^-2·d^-1 with supplemental lighting (Figure 3), closely matching the mean DLI during the data collection.

Diurnal patterns of environmental conditions were visualized by averaging values at each hour of day across the data collection period (Figure 4). PPFD exhibited a largely symmetric, near-sinusoidal pattern, with a midpoint at approximately 13:30, corresponding to solar noon at the experimental location. A localized reduction in PPFD was observed around 13:00, likely due to transient shading from the greenhouse structure over the measurement area. Temp and VPD increased progressively during the photoperiod, whereas CO₂ declined. Consistent with these asymmetric diurnal patterns in environmental variables, except for light intensities, Φ_PSII displayed a modest post-noon decline. The average Φ_PSII was 0.658 before 13:30 and 0.624 after 13:30.

Correlation analysis revealed strong positive relationships among light intensity-related variables (ePPFDi, ePPFD15, PPFDi, and PPFD15) and among cumulative light variables (eDLI_cum and DLI_cum) (Figure 5). Temp and VPD were also positively correlated with each other and negatively correlated with CO₂, reflecting their opposing diurnal trends (Figure 4). Since cumulative light variables increased over time, DLI_cum and eDLI_cum showed positive relationships with Temp and VPD and a negative relationship with CO₂ (Figure 5).

Bivariate relationships between individual environmental predictors and Φ_PSII were analyzed independently to provide an initial assessment of relative influence prior to multivariate modeling (Table 1). Light intensity-related variables exhibited the highest R² values, followed by CO₂, Temp, and VPD, whereas cumulated light variables showed the lowest R². Across both ePPFD- and PPFD-based metrics, previous 15-min averaged light intensity variables (ePPFD15 and PPFD15) consistently explained more variation in Φ_PSII than instantaneous measurements (ePPFDi and PPFDi). In addition, light intensity metrics that included far-red radiation (ePPFD15 and ePPFDi) were associated with higher R² values than their PPFD-based variables (PPFD15 and PPFDi).

Overall, the final MLR model integrated short-term light environments, atmospheric conditions, and a temporal predictor to predict Φ_PSII. The selected predictors included instantaneous and 15-min-averaged ePPFD (ePPFDi and ePPFD15), along with Temp, VPD, and CO₂. A binary Time-of-day variable was retained in the final model to account for the diurnal asymmetry of Φ_PSII. In contrast, PPFD- and DLI-based variables were not retained in the final model.

Model performance was evaluated using a training-testing data split (Table 3). The model exhibited comparable predictive performance between the training and testing datasets, with RMSE values of 0.083 and 0.080 and R² values of 0.56 and 0.60, respectively. Predicted Φ_PSII values followed measured values, with observations distributed around 1:1 line and no strong systematic bias across the prediction range (Figure 6). Diagnostic evaluation indicated no substantial violations of model assumptions, including linearity, residual normality, and multicollinearity (Supplementary Figures 2, 3; Supplementary Table 1). Temporal autocorrelation was not detected based on the Durbin-Watson test (DW = 1.999, P = 0.491).Variable importance analysis indicated that light intensity-related predictors contributed most strongly to model predictions (Figure 2). Among these, ePPFD15 had the highest importance, followed by ePPFDi. VPD and CO₂ showed intermediate importance, whereas Temp had the lowest importance among the continuous predictors.

Regression coefficients revealed negative associations between Φ_PSII and ePPFD15, ePPFDi, and Temp, whereas positive coefficients were observed for VPD and CO₂ (Table 2). The positive coefficient of the Time-of-day indicates that Φ_PSII tended to be higher during the beforenoon compared to the afternoon.

Both sensor-based and ML-based control strategies maintained the target ETR of 100 μmol·m^-2·s^-1 throughout the experimental period, with average ETR values of 100.9 ± 11.1 (measured) for BFB-ETR 100 and 101.6 ± 8.7 μmol·m^-2·s^-1 (predicted) for ML-ETR 100 (Table 4).

For the Φ_PSII-based strategies, the ML-Φ_PSII 0.68 achieved an average predicted Φ_PSII of 0.678 ± 0.018, closely matching the target value (Table 4). In contrast, BFB-Φ_PSII 0.68 resulted in a lower measured Φ_PSII of 0.646 ± 0.039 and was associated with lower average PPFD and DLI over the experimental period (Figure 3). In all other treatments, average PPFD and DLI were comparable at approximately 350 μmol·m^-2·s^-1 and 16 mol·m^-2·day^-1, respectively.

Figure 7 illustrates diurnal patterns of supplemental light adjustment within the photoperiod. The PPFD 350 treatment maintained a stable PPFD close to 350 μmol·m^-2·s^-1 throughout the photoperiod, with brief deviations during midday solar peaks (Figure 7E). In contrast, both BFB-ETR 100 and ML-ETR 100 exhibited slightly greater within-day variability in PPFD relative to constant PPFD treatment (Figures 7A, C), while ETR values were maintained near the target level of 100 μmol·m^-2·s^-1 (Figures 8A, C). In the ML-ETR 100 treatment, a transient increase in PPFD during the morning hours coincided with a pronounced decline in predicted Φ_PSII (Figures 7C, 9C). By comparison, BFB-ETR 100 treatment showed an increase in measured Φ_PSII during the mid-photoperiod, followed by stabilization later in the day (Figure 9A).

Both BFB- and ML-Φ_PSII 0.68 treatments exhibited greater within-day fluctuations in supplemental light regulation compared with the ETR-based strategies and the constant PPFD treatment (Figures 7B, D). These fluctuations were also reflected in corresponding diurnal patterns of ETR (Figures 8B, D). Despite the dynamic light adjustments, measured and predicted Φ_PSII values in the Φ_PSII-based treatments remained relatively stable throughout the photoperiod (Figures 9B, D). However, under the PPFD 350 treatment, predicted Φ_PSII exhibited pronounced diurnal variation throughout the photoperiod, compared with four biofeedback-based strategies (Figure 9E).

Shoot fresh weight, shoot dry weight, leaf chlorophyll content, and leaf anthocyanin content did not differ significantly among treatments (Table 5). Total supplemental light energy use varied among treatments, with ML-Φ_PSII 0.68 exhibiting the highest cumulative energy consumption, followed by PPFD 350 and ML-ETR 100. In contrast, sensor-based biofeedback treatments showed lower total energy use. Energy use efficiency, expressed as biomass production per unit of supplemental light energy, differed significantly among treatments. BFB-ETR 100 showed the highest energy use efficiency based on both fresh and dry weight, whereas PPFD 350 had the lowest values. Based on fresh weight, BFB-Φ_PSII 0.68 produced lower energy use efficiency than BFB-ETR 100, but the remaining treatments showed intermediate values with no significant differences.