Two Post-1995 Scientists Link AI and Materials Science in a Closed-Loop Lab

Xu and An are young materials science scientists at the institute and algorithm experts at Gewu Zhiyan, a company incubated by the institute. One came from chemistry into AI; the other from physics to computer science and then to AI for Science. The live broadcast left little room for error. Xu's tensest moment was ligand recommendation, because ligands are a core part of catalysts and determine reaction efficiency and selectivity. He did not know whether the model's recommended ligands would raise yields and could not run extra verification during the broadcast. 'I was very lucky,' Xu recalled. 'It did improve the average yield of four reactions, and I did not know this experimental result beforehand.' An, responsible for the Agent system, was most worried about drug discovery. Those tasks involve long chains and many tool calls, with a single run lasting five or even ten hours; errors can accumulate and the Agent can collapse midway. 'During the livestream, drug discovery achieved an unprecedented task success rate,' An said.

Xu studied applied chemistry at East China University of Science and Technology. In his sophomore year, as AlphaGo drew attention, he taught himself Python and later joined a computational chemistry group at Zhejiang University led by Hong Xin, working on machine learning for chemistry when the direction was still unusual. His first task was collecting wet-lab reaction data from literature. Over about 18 months, he gathered more than 10,000 data points on olefin asymmetric hydrogenation from hundreds of papers, and during the pandemic added several thousand more. The dataset remains the largest structured asymmetric catalytic reaction dataset and was published in Angewandte Chemie. After a brief ByteDance internship, he took part in the 2023 World AI Science Challenge hosted by Fudan University and the institute, and then joined the institute.

An entered Nanjing University's physics department but found the discipline increasingly harsh. He switched to computer science in 2016, as Kaiming He's ResNet opened a new deep learning era, and studied under Professor Shen Furao. Shen guided him to integrate energy minimization and Hopfield network models into neural networks. An's PhD began with image classification, segmentation and detection, then moved to 3D point clouds, where he saw what he called the 'beauty of symmetry' in rotation equivariance formulas. He worked on point cloud registration and won first prize in an ICCV competition. In 2023, Qu Chao, now an AI scientist at the institute, told him that molecules and atoms are in a sense 3D point clouds, but more 'obedient' than LiDAR points: they have no noise and fixed information, though they require far more complex and certain physical priors. Through Qu, An met Cao Fenglei, the institute's chief scientist for materials science intelligence, who described a blueprint for a 3D chemical space representation model. An joined after his PhD, with Cao later as his direct leader.

Xu and An joined in early 2024. Xu first proposed an Agent system, but the institute's dean, Qi Yuan, did not approve it, saying the team lacked model accumulation and should first build a technically demanding model. Xu then developed a general reaction prediction model for yield, selectivity and synthesis route planning. The work was published in Nature Machine Intelligence and appeared on the cover. Its core innovation was a reaction-oriented pretraining strategy about how chemical bonds form and break. Some classmates and reviewers suggested the common NLP next token prediction approach, but Xu insisted on his own. He then started transition-state prediction. In a reaction, molecules must overcome an energy barrier, like crossing a small hill. Searching for transition states is extremely difficult because algorithms usually drive energy downward while transition states require going upward, and chemists must build initial structures very close to the real one. Xu wanted an end-to-end method: draw a 2D transition-state structure and directly generate the 3D version. He struggled for eight months with chaotic results. 'If my first work at the institute had been transition-state generation and I had gone through more than eight months of failure, I definitely could not have held on,' he recalled. 'But my first project went smoothly and at least proved myself.' In autumn 2024, he introduced higher-order equivariant graph neural networks into diffusion denoising models, a term An repeatedly mentioned. An had used the technique in EST, the Equivariant Spherical Transformer, a core module of the Suiren model. Xu's attempt worked; he continued to increase the order and completed the core module. As part of the Suiren materials science model series, the work was published as UniTS in Nature Communications at the end of August and was used in the Super Research Factory.

An's task was to build a general representation model for chemical space. By then, the emergent abilities of large models were clear, and An wondered whether molecular representation models could show similar emergence. This became the starting point for the base layer of the Suiren series. In discussions with Qi Yuan, Cao Fenglei and Xu, Qi asked whether scaling law appears in the molecular domain; the team listed factors from parameter count to compute and discussed them one by one. An had hesitated because training large models is costly, but after the discussion he decided to advance once a route looked good in theory and intuition. Training went relatively smoothly, but evaluation and application did not open a clear gap over competitors. Many chemical and physical downstream tasks depend on macroscopic molecular behavior, and modeling the Boltzmann distribution is a problem that has stalled most researchers worldwide. An tried various structures to compress microscopic representations into macroscopic ones, with poor results. In late 2025, he realized that diffusion algorithms could perform gradual conformation compression, with macroscopic representations as conditions; he had previously treated diffusion only as a generative model. He introduced the algorithm, and early 2026 evaluations showed the model could directly solve macroscopic tasks and had made a leap over competitors. 'I did not fall asleep until three or four in the morning that night, purely from excitement,' he said.

Both had crossed their barriers, but their work needed a larger container. The Suiren model needed reaction prediction and transition-state generation; dry and wet experiments needed an Agent to connect them; and the Agent needed a laboratory for validation. At the end of 2025, the Huntianling Scientific Skills Platform was launched. Xu described An's work as the 'base': molecules and atoms are the microscopic units of the material world, and understanding their representations and mechanisms supports higher-level applications. An described Xu's work as the 'bridge': Xu understands what molecules can do, but whether and how molecules can be obtained in the physical world, and what reaction mechanisms they follow, 'must depend on Licheng's organic synthesis and catalysis work.' An cared more about algorithm design; Xu cared more about whether physical chemistry mechanisms were correct and explainable. They called the collaboration 'high cohesion, low coupling': each person is an independent module with defined inputs and outputs. In practice, An said, they would 'understand one layer above,' knowing not just what the other wanted but why. Huntianling was suggested by Cao Fenglei. The team had developed many models and tools, but they were too difficult for people without computational chemistry knowledge to use directly; an Agent was a good vehicle to connect research workflows. The platform defined only batteries, drugs, reactions and materials workflows. An wrote the Agent's underlying code and framework; Xu handled encapsulation of chemical tools and integration of dry and wet experiment modules. In one example, Xu wrote scripts under chemists' guidance to read reaction conversion from LC-MS spectra, and An merged the scripts into the workflow so experimental results could match natural language. Dry experiments are not the endpoint. Materials science must translate virtual plans into the physical world, and the laboratory platform connects the dry-wet loop. In March 2026, the self-driving laboratory of the Golab Materials Science Intelligent R&D Factory was formally approved. In early construction, An connected the most difficult link, drug discovery, which took about half a month to get running.

Editor's Summary

Leiphone's report profiles Xu Licheng and An Junyi, two post-1995 scientists at the Shanghai Institute for AI for Science, whose work spans reaction prediction, transition-state generation, Suiren models and the Golab dry-wet laboratory. Their efforts were showcased in a 120-hour, 135-topic livestream and contributed to the Huntianling platform and a self-driving laboratory project approved in March 2026. The account illustrates how young researchers are combining AI with chemistry and physics to build closed-loop materials science tools.