Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 150056

150056 - A Digital Twin Model for Sorghum Growth Simulation and Breeding Optimization 

Given the future demand for food, especially under a changing climate, there is a need for sophisticated, broad-scale, and genetically informed breeding approaches for the improvement of our crops. One of the most versatile and resilient crops to increase stress from a changing climate is sorghum, a natural and heat-tolerant plant that can withstand high temperatures and a variety of soil types. An intelligent digital twin that can predict crop growth will help us improve breeding decisions and increase yields in an agile and sustainable way. In response to the necessity to breed crops faster and produce more and with fewer inputs compared with conventional breeding, we have developed a digital twin model for sorghum growth that introduces Genotype x Environment x Management (GxExM) interactions into breeding pipelines.

Our complement to the field of agricultural science was creating a bi-stage digital twin model that traverses from the genotype to the phenotype through "trenotype", translated genotype. The digital twin considers the spatiotemporal weather data, soil condition, and pedigree information, thus providing a powerful tool for heritable analysis of the hybrids' genomic traits. The digital twin model is supported by a massive data set containing the hourly weather recordings for seven years, soil data collected for five depths, and the pedigree data of 665 fathers and 141 mothers. With the help of the multi-modal dataset, a given management practice is simulated sufficiently for the prediction of its phenotypic outcomes.

We developed a bi-stage model to fit the empirical data, where the first stage converts the genotype data into translated genetic properties, "trenotype", accounting for additive, dominant, and epistatic effects, and the second stage predicts phenotypes by compiling trenotype data with environmental and management factors to yield outputs such as daily biomass accumulation and final or harvest yield. We evaluated the model's performance comprehensively through a Relative Root Mean Squared Error (RRMSE) criterion, which ranged from 8 to 10 percent, a value that signifies a high level of accuracy and reliability.

Preliminary results from the practical application of our model in breeding programs have shown significant improvements. Specifically, the digital twin model achieved a 9% increase in the maximum recovery rate of high-yielding hybrids within just two generations compared to traditional breeding methods. This result highlights the model's efficiency in identifying superior hybrids, thereby accelerating the breeding process and reducing the time and resources required. The active learning framework, which iteratively refines predictions based on new data, ensures that the model remains adaptable and accurate in dynamic genetic and environmental landscapes.

In conclusion, our digital twin model represents a significant advancement in crop breeding technology. By providing accurate and efficient predictions of phenotypic outcomes, it supports the development of resilient, high-yielding sorghum varieties, contributing to sustainable agricultural practices and food security. The model's continuous learning and adaptation capabilities make it a valuable tool for ongoing breeding efforts. Future research should focus on expanding the model to include heterozygous parents and multiple generations, as well as adapting it for other crops to enhance its applicability and impact in agricultural science.

Presenting Author: Zheng Ni Oklahoma State University

Presenting Author Biography: Zheng Ni is a Ph.D. student in industrial engineering at Oklahoma State University advised by Prof. Lizhi Wang. Currently, he is doing research in plant breeding and yield prediction using crop modeling, machine learning, and Monte Carlo simulation.

Authors:

Zheng Ni Oklahoma State University
Yanbin Chang Oklahoma state university
Maria Salas Fernandez Iowa state university
Lizhi Wang Oklahoma state university