PhoTorch

PhoTorch is a robust and generalized photosynthesis biochemical model fitting package based on PyTorch.

To cite this package, please use:

Lei, T., Rizzo, K. T., & Bailey, B. N. (2025). PhoTorch: a robust and generalized biochemical photosynthesis model fitting package based on PyTorch. Photosynthesis Research, 163(2), 21. https://doi.org/10.1007/s11120-025-01136-7.

Note: The latest version 1.3.0 includes changes to the file structure and function names within the package.

Currently, the package includes the Farquhar, von Caemmerer, and Berry (FvCB) model and stomatal conductance models including Buckley Mott Farquhar (BMF), Medlyn (MED), and Ball Woodrow Berry (BWB) versions. The Ball Berry Leuning (BBL) stomatal conductance model and the PROSPECT models are under development.

Installation of dependencies

pip install torch
pip install numpy
pip install scipy
pip install pandas
pip install matplotlib

Next, download the repository, and try running the examples in the testphotorch.py file.

Repository download

git clone https://github.com/GEMINI-Breeding/photorch.git

or

pip install photorch

1. FvCB model usage

Create a python file in the PhoTorch directory and import necessary packages.

from photorch import *
import pandas as pd
import torch

Load data

Load the example CSV file. The loaded data frame should have columns with titles 'CurveID', 'FittingGroup', 'Ci', 'A', 'Qin', and 'Tleaf'. Each A/Ci curve should have a unique 'CurveID'. If no 'Qin' and 'Tleaf' are available, it will be automatically set to 2000 and 25, respectively.

The data to be loaded should be:

| CurveID | FittingGroup | Ci | A | Qin | Tleaf | |---------|--------------|-----|----|------|-------| | 1 | 1 | 200 | 20 | 2000 | 25 | | 1 | 1 | 400 | 30 | 2000 | 25 | | 1 | 1 | 600 | 40 | 2000 | 25 | | 2 | 1 | 200 | 25 | 2000 | 30 | | 2 | 1 | 400 | 35 | 2000 | 30 | | 2 | 1 | 700 | 55 | 2000 | 30 |

dftest = pd.read_csv('data/dfMAGIC043_lr.csv')

Initialize the data

Then, specify the ID of the light response curve. If there is no light response curve in the dataset, ignore it (default is None).

# Specify the list of light response curve IDs, if no light response curve, input "lightresp_id = None" or ignore it.
lcd = fvcb.initLicordata(dftest, preprocess=True, lightresp_id = [118])

Define the device

Default device is 'cpu'. If you have an NVIDIA GPU, set 'device_fit' to 'cuda' and execute the 'lcd.todevice(torch.device(device_fit))' line.

device_fit = 'cpu'
lcd.todevice(torch.device(device_fit)) # if device is cuda, then execute this line

Initialize FvCB model

If 'onefit' is set to 'True', all curves in a fitting group will share the same set of Vcmax25, Jmax25, TPU25, and Rd25. Otherwise, each curve will have its own set of these four main parameters but share the same light and temperature response parameters for the fitting group.

If no light response curve is specified, set 'LightResp_type' to 0.

LightResp_type 0: J is equal to Jmax.

LightResp_type 1: using equation $J = \frac{\alpha Q J_{max}}{\alpha Q + J_{max}}$ and fitting $\alpha$.

LightResp_type 2: using equation $J = \frac{\alpha Q + J_{max} - \sqrt{(\alpha Q + J_{max})^2 - 4 \theta \alpha Q J_{max}}}{2 \theta}$ and fitting $\alpha$ and $\theta$.

TempResp_type 0: Vcmax, Jmax, TPU, and Rd are equal to the Vcmax25, Jmax25, TPU25, and Rd25, respectively.

TempResp_type 1: using equation $k = k_{25} \exp{\left[\frac{\Delta{H_a}}{R}\left(\frac{1}{298}-\frac{1}{T_{leaf}}\right)\right]}$ and fitting $\Delta{H_a}$ for Vcmax, Jmax, and TPU.

TempResp_type 2: using equation $k = k_{25} \exp\left[\frac{\Delta H_a}{R} \left(\frac{1}{298}-\frac{1}{T_{leaf}}\right)\right] \frac{f\left(298\right)}{f\left(T_{leaf}\right)}$, where $f(T) = 1+\exp \left[\frac{\Delta H_d}{R}\left(\frac{1}{T_{opt}}-\frac{1}{T} \right)-\ln \left(\frac{\Delta H_d}{\Delta H_a}-1 \right) \right]$, and fitting $\Delta{H_a}$ and $T_{opt}$ for Vcmax, Jmax, and TPU.

# initialize the model
fvcbm = fvcb.model(lcd, LightResp_type = 2, TempResp_type = 2, onefit = False)

More fitting options

fitRd: option to fit $R_{d25}$, default is True. If set to False, $R_{d}$ will be fixed to 1% of $V_{cmax}$.

fitRdratio: option to fit $R_{d}$-to- $V_{cmax}$ ratio, default is False, the range is 0.01 to 0.02.

fitag: option to fit $\alpha_g$, default is False, the range is 0 to 1.

fitKc: option to fit $k_{25}$, default is False.

fitKo: option to fit $k_{25}$, default is False.

fitgamma: option to fit $\Gamma^*_{25}$, default is False.

fitgm: option to fit $g_m$, default is False.

fvcbm = fvcb.model(lcd, LightResp_type = 0, TempResp_type = 1, onefit = False, fitRd = True, fitRdratio = False, fitag = False, fitgm= False, fitgamma=False, fitKo=False, fitKc=False, allparams=allparams)

Specify default fixed or learnable parameters

allparams = fvcb.allparameters()
allparams.dHa_Vcmax = torch.tensor(40.0).to(device_fit) # If the device is cuda, then execute ".to(device_fit)"
fvcbm = fvcb.model(lcd, LightResp_type = 0, TempResp_type = 1, onefit = False, fitag = False, fitgm= False, fitgamma=False, fitKo=False, fitKc=False, allparams=allparams)

Fit A/Ci curves

fitresult = fvcb.fit(fvcbm, learn_rate= 0.08, maxiteration = 20000, minloss= 1, recordweightsTF=False)
fvcbm = fitresult.model

Get fitted parameters by ID

The main parameters are stored in the 'fvcbm'. The temperature response parameters are in 'fvcbm.TempResponse', just like the light response parameters.

id_index = 0
id = int(lcd.IDs[id_index]) # target curve ID
fg_index =  int(lcd.FGs_idx[id_index]) # index of the corresponding fitting group
if not fvcbm.onefit:
    Vcmax25_id = fvcbm.Vcmax25[id_index]
    Jmax25_id = fvcbm.Jmax25[id_index]
else:
    Vcmax25_id = fvcbm.Vcmax25[fg_index]
    Jmax25_id = fvcbm.Jmax25[fg_index]

dHa_Vcmax_id = fvcbm.TempResponse.dHa_Vcmax[fg_index]
alpha_id = fvcbm.LightResponse.alpha[fg_index]

Get fitted A/Ci curves

A, Ac, Aj, Ap = fvcbm()

Get fitted A/Ci curves by ID

id_index = 0
id = lcd.IDs[id_index]
indices_id = lcd.getIndicesbyID(id)
A_id = A[indices_id]
Ac_id = Ac[indices_id]

Get the (preprocessed) photosynthesis data by ID

A_id_mea, Ci_id, Q_id, Tlf_id = lcd.getDatabyID(lcd.IDs[id_index])

Routine Single Group Fitting, Plotting, and Evaluating

For routine fitting of single fitting groups, an easy to use example is available.

From here, one can easily change model settings in the User Settings section of each code block. The first block will process your data (compile all response curves) and the second block with fit the processed data with PhoTorch.

# Process ACi Data and Verify the Printed Curves are Those Desired to Fit

################# User Settings #################

fitting_group_folder_path = "photorch/data/fvcb/curves/iceberg"
species_to_fit = "Iceberg"
species_variety = "Calmar"

#################################################
from photorch.src.util import *

compiledDataPath = compileACiFiles(fitting_group_folder_path)

# Fit Compiled Data to FvCB Model using PhoTorch, Save Parameters, and Plot Results

######### User Settings ##########
LightResponseType = 2
TemperatureResponseType = 2
Fitgm = False
FitGamma = False
FitKc = False
FitKo = False
saveParameters = True
plotResultingFit = True
#### Advanced Hyper Parameters ####
learningRate = 0.08
iterations = 10000
###################################

The file structure is used in the processing and plotting pipeline so familiarize yourself with it. Data for the FvCB model fitting is housed under "data/fvcb/curves/{FittingGroup}" for the response curves, and "data/fvcb/survey{FittingGroup}" for any auxillary survey measurements. E.g.

data/fvcb/curves/iceberg/

The {FittingGroup}s are species, varieties, or any collection of data used to fit one set of parameters with. The {FittingGroup} folder expects raw LI-6800 ASCII files (without extensions) that are output alongside the optional .xlsx files (which pandas currently cannot import). The processing step converts these raw files to .txt before extracting and compiling the curves. The species_to_fit (required) and species_variety (can be left as an empty string, $``"$) variables are just to name the saved parameters and plots, and are not required to match the path given.

Parameters will optionally be saved to

results/parameters/{species_to_fit}{species_variety}_FvCB_Parameters.csv

and figures will optionally be saved to

results/figures/{species_to_fit}{species_variety}_FvCB_{figure_name}.png

The normalization of response curves from different leaves using survey data is currently under development.

---

2. Stomatal conductance model usage

Three stomatal conductance model is currently available: Buckley Mott Farquhar (BMF), Medlyn (MED), and Ball Woodrow Berry (BWB). The Ball Berry Leuning (BBL) model is under development. More details about these four models can be found at: https://baileylab.ucdavis.edu/software/helios/_stomatal_doc.html.

Create a python file in the PhoTorch directory and import necessary packages.

from photorch import stomatal
import pandas as pd
import torch

Initialize the stomatal conductance data

The data to be loaded should be:

| CurveID | gsw | VPDleaf | A | Qin | Tleaf | RHcham | |---------|-----|---------|----|-------|-------|--------| | 0 | 0.34| 11.32 | 55 | 2000 | 21.81 | 21.0 | | 0 | 0.34| 18.33 | 55 | 2000 | 22.71 | 30.02 | | 0 | 0.35| 29.57 | 51 | 2000 | 20.02 | 38.01 | | 0 | 0.38| 15.4 | 54 | 2000 | 22.6 | 26.99 | | 0 | 0.32| 15.44 | 54 | 1200 | 19.97 | 27.0 | | 1 | 0.23| 29.