“Agentic AI will become a normal tool in scientific research”: Meet the scientist enabling the labs, materials and manufacturing of tomorrow

Dr Tong Xie will give a keynote address to open day two of ACM CRC’s Partner Meeting on August 20. Here, the Associate Lecturer at UNSW Sydney’s School of Photovoltaic and Renewable Energy Engineering (SPREE) shares some insights about AI for “knowledge extraction, hypothesis generation, multimodal modelling and autonomous experimentation”, and on turning scientific discoveries into impact outside of the lab.      

ACM CRC: What does your research/commercial career look like so far, and what are some things that have shaped it?

Dr Tong Xie: My career has developed at the intersection of materials science, renewable energy and artificial intelligence.

During my PhD at UNSW, I became interested in why scientific discovery remains so slow and fragmented. Researchers still spend enormous amounts of time reading papers, developing hypotheses, running experiments and repeating the process.

This led me to work on scientific foundation models and AI agents, including DARWIN and MOOSE-Chem. I later founded GreenDynamics to translate that research into practical systems for laboratories and industry.

What has shaped me most is seeing both sides: academic research needs scientific depth, while commercial deployment requires reliability, integration and clear real-world value.

ACM CRC: What did your PhD involve?

TX: My PhD focused on using large language models to accelerate nano material discovery and synthesis.

I studied how nanomaterial composition, processing conditions and device structures affect solar-cell performance and stability. These are highly complex problems, where small changes in solvents, temperatures or processing steps can significantly affect the final result.

That experience led me to explore whether AI could combine scientific knowledge with first -principle theory and experimental data to help researchers generate better hypotheses and design more effective experiments.

ACM CRC: You’re at UNSW’s famous SPREE. How do AI and solar research overlap in your current work?

TX: Solar-cell development is fundamentally a materials-discovery and manufacturing-optimisation problem.

A solar device contains many different materials, interfaces and processing conditions, creating a very large experimental search space. AI can help identify promising materials, predict properties, recommend experiments and analyse why a device succeeds or fails.

ACM CRC: Where does your entrepreneurial streak come from? And can researchers learn a commercial mindset?

TX: I have always been interested in building things that people can actually use.

My entrepreneurial motivation came partly from seeing how difficult it was to move strong academic AI results into real laboratories and manufacturing environments. Industry needs systems that are reliable, integrated and economically valuable, not only models that perform well on benchmarks.

I believe researchers can benefit from commercial thinking. It encourages them to ask who has the problem, how important it is, what evidence is required and how a solution would fit into an existing workflow.

Commercial thinking should not replace scientific curiosity, but it can help turn research into impact.

ACM CRC: Talk us through the steps in discovery and how LLMs could help.

TX: Scientific discovery usually begins with understanding existing knowledge. LLMs can review large volumes of literature, identify connections and highlight gaps or disagreements.

They can then help generate hypotheses, propose materials and recommend experiments. After an experiment, AI can analyse data such as images, spectra, compositions and process records.

The most important stage is iteration. When the model is connected to an automated laboratory, it can propose an experiment, observe the result, update its reasoning and recommend the next step.

The goal is not simply an AI that answers scientific questions, but one that supports the full discovery cycle while keeping researchers in control.

ACM CRC: What else are you applying LLMs and agentic AI to at UNSW?

TX: We work across scientific knowledge extraction, hypothesis generation, multimodal modelling and autonomous experimentation.

Our models combine scientific text with chemical compositions, crystal structures, numerical properties and experimental information.

Through projects such as AtomWorld, we are also studying whether AI models can take reliable scientific actions. A model may be able to describe an atomic structure but still struggle to manipulate it correctly.

This is an important distinction: understanding scientific language does not automatically mean a model can operate reliably in the physical world.

ACM CRC: How and why did you establish GreenDynamics?

TX: I established GreenDynamics because I saw a gap between scientific AI research and the tools available to industrial researchers.

Many AI systems stop at prediction. They may recommend a material, but the customer still needs to determine whether it can be synthesised, tested, manufactured and scaled.

GreenDynamics was created to close that loop. ByteScience supports scientific reasoning and decision-making, while ByteFactory connects AI with physical laboratory workflows.

Our aim is to move from a discovery objective to experimentally validated materials and processes.

ACM CRC: What is its focus, and how does its business model look?

TX: GreenDynamics focuses on AI-driven development of advanced materials, including formulations, coatings, functional films, particles and surface-processing systems.

We currently work with industrial partners on specific materials and process challenges. Our business model includes AI software and longer-term licensing or joint development. The ultimate goal is to build a material IP factory.

ACM CRC: How do you plan to scale up?

TX: In the near term, we are focusing on selected industrial applications where the experimental variables are clear and the commercial value is significant.

We are also standardising the connection between ByteScience and ByteFactory so that our systems can be adapted across different materials and industries.

In the medium term, we aim to build a distributed network of AI-connected laboratories. This would allow us to support the full journey from identifying a scientific opportunity to validation and manufacturing scale-up.

ACM CRC: How will agentic AI benefit advanced-material discovery and manufacturing?

TX: I believe agentic AI will become a normal tool in scientific research.

Initially, agents will support literature review, knowledge organisation and reporting. They will then move towards recommending experiments and analysing complex data.

The biggest change will come when agents are connected to laboratories and manufacturing systems, allowing them to learn from physical results and adjust the next action.

However, the standard for reliability is much higher when AI moves from producing text to controlling experiments. These systems must understand uncertainty, provide traceable evidence and know when human judgement is required.

I often describe this as action scaling. Progress will not come only from larger models, but from giving AI access to real experimental feedback and teaching it to act reliably in the physical world.

ACM CRC: Any closing words about your Partner Conference speech?

TX: My presentation will cover the evolution from language models that can read science, to agents that can reason, and ultimately to AI systems that can interact with laboratories.

I will share examples from our work in scientific foundation models, photovoltaics, hypothesis generation and autonomous experimentation.

My central message is that AI for science must be developed together with scientists, laboratories and manufacturers.

The future laboratory will not be operated by AI alone. It will be a collaboration between researchers, AI agents and automated instruments.

ACM CRC’s annual Partner Meeting will be held on August 19 and 20 at Level 3, Salesforce Tower, Sydney. Dr Tong Xie’s keynote will open day two at this invite-only event. You can read more about Xie’s research at his UNSW staff profile page, and more about his venture, GreenDynamics, here.