AI cracks the code for faster, better crops

Connecting data islands

For generations, breeders have operated like explorers in a vast, uncharted biological wilderness. The process of selecting parent plants, crossbreeding, and evaluating thousands of progeny over multiple growing seasons is painstakingly slow, with the success rate often below 1 percent, experts said.

This challenge, Chen added, is compounded by deeply entrenched "data silos".

"Data on genotype, phenotype, environment, and even soil are all kept separate. This fragmentation creates a critical bottleneck," he said.

Yuan said: "Researchers often know neither the source and quality of data, nor can they discern which data AI can understand. This causes AI to falter — or worse, produce erroneous results."

It is this precise problem that the Fan project is engineered to solve, acting as a "central nervous system" to connect disparate data islands — a full-chain AI technical system built on Huawei's AI data solution.

Yuan Yuan, vice-president of Huawei's data storage product line, said the Fan platform tackles the problem in three ways. First, it aggregates and standardizes multisource data on environment, traits, phenotype, and genotype from across the country.

Second, it utilizes specialized tools to enable the rapid construction of customized, industry-specific AI large language models, which can cut model development time from 15 days to five, Yuan said.

Finally, its core "breeding AI agent" can intelligently screen this unified data, automate complex analysis workflows, and validate models to identify optimal breeding pathways, he said.

"The impact is transformative," Yuan said.

"It can shorten the traditional 20-generation cultivation cycle for crops like rice, which usually takes eight to 10 years, to just five generations, or three to four years."

This represents a 50 percent reduction in the breeding cycle and can boost overall efficiency by an estimated 30 percent.

The project represents more than a technical advancement. It is also a statement of strategic intent aligned with a national blueprint. "This intelligent system currently does not exist globally," said Chen.

The goal is to rapidly advance the construction of the "Nanfan Silicon Valley" and establish a leading hub for future agriculture.

"Nanfan" refers to a unique breeding method using Hainan's warm winters as a natural way to accelerate the process. According to a national plan, the Nanfan breeding base, located in Hainan, is set to evolve into the "Silicon Valley" of China's seed industry by 2030, serving as a comprehensive hub for agricultural research, industry, and technology exchange.

This ambition mirrors high-level national directives. On Nov 13, 2025,China's Ministry of Agriculture and Rural Affairs convened a national conference to advance the seed industry revitalization action, charting the course for the 15th Five-Year Plan period (2026-30). The conference called for accelerating the realization of self-reliance and self-improvement in seed technology and securing a firm grip on seed sources.

At the industry level, the plan emphasizes upgrading the Nanfan Silicon Valley scientific base into a national seed innovation hub that integrates research, commercialization, and application.

"Digitalization and intelligence are undoubtedly the future directions for building the Nanfan Silicon Valley," Chen said. "We must use advanced technology to serve and transform both agricultural production and research."

This initiative is part of a broader push to harness AI for agricultural progress across the nation. In 2024,Yazhou Bay National Laboratory researchers, in collaboration with China Agricultural University and the Shanghai Artificial Intelligence Laboratory, developed China's first large language model for seed design, known as SeedLLM, or Fengdeng.

This AI platform provides expert insights on breeding, cultivation and industry trends — empowering farmers and researchers with practical knowledge.

In July 2025, Fengdeng was upgraded to an AI agent with three core research functions, said Yang Fan, a scientist at the laboratory.

The first function is knowledge summarization, which addresses key questions like "which traits are regulated by what type of genes". It does this by automatically integrating over 98 percent of relevant global crop research literature to build a gene-trait-environment association map.

The second is gene-trait association prediction, enabling autonomous genome-wide screening of key genes beyond traditional reasoning.

The third is experimental reasoning and design optimization, where it simulates expert logic to automate the entire research cycle from hypothesis generation and experimental design to result analysis, Yang said.