How Singapore Is Using AI to Turn Scientific Discoveries Into Real-World Technologies

Singapore is treating AI for science as more than a faster way to search papers or predict molecules. Its National Research Foundation (NRF) has committed S$120 million to eight projects intended to place AI inside the scientific workflow—from forming hypotheses and selecting experiments to analysing results, verifying software and developing technologies that can be tested outside the lab. The initiative was formally launched on 16 June 2026.

The important distinction is between accelerating discovery and proving deployment. The projects are designed to close that gap, but their promised materials, catalysts, diagnostics and software are still research objectives rather than established commercial products. Several projects are expected to run for four to five years.

The Materials Data Foundry tackles AI’s manufacturing problem

AI can identify a material that appears promising in a simulation. That does not necessarily reveal how to make it reliably, at scale or with the performance predicted by the model. The S$10 million Materials Data Foundry, jointly established by the National University of Singapore and the University of Toronto’s Acceleration Consortium, is designed around this prediction-to-production bottleneck.

The planned facility will combine AI, robotics and in-situ measurement in an open autonomous laboratory. Rather than relying only on published materials data, it will generate experimental data linking synthesis conditions to real-world material performance. Those measurements can then improve the models and guide the next experiments, creating a feedback loop between prediction and physical testing.

The project’s stated targets include large numbers of experiments and practical, reproducible synthesis recipes for areas such as beyond-silicon or quantum materials, electrocatalysts relevant to clean hydrogen and corrosion-resistant coatings. The intended output is therefore not simply a list of AI-generated candidates; it is evidence about how promising materials can actually be produced and evaluated.

AI will search catalyst designs before researchers test them

A separate S$10 million programme led by A*STAR and Imperial will develop a Surface Science Foundation Model and a cloud platform for predicting catalyst properties and synthesis conditions. The research is aimed at applications including cleaner fuels, carbon-dioxide-derived feedstocks and bio-based chemicals.

The programme description cites a target of 500-fold acceleration in computational screening. That is a screening objective, not proof that a finished fuel or chemical process will arrive 500 times faster. Claims that the work will be “nearly five times faster and cheaper” should be treated cautiously unless they refer to a separate downstream metric. The decisive test will be whether computationally selected catalysts perform well in experiments and can be developed into viable processes.

Software projects apply scientific verification to AI-generated code

AI4S also treats software reliability as a scientific and engineering research problem. NUS and Imperial are developing program-reasoning tools that can analyse and test code, formally verify properties and prove correctness where possible. The tools are intended for difficult or consequential software, including network protocols and Linux-kernel components.

One proposed use is to let AI systems audit code produced by other AI systems. That could make AI-assisted programming more dependable, but automated analysis is not the same as a universal guarantee of safety. Formal methods can establish specific properties under defined assumptions; broader reliability still depends on the quality of the specifications, tests and systems in which the code operates.

Health projects combine biological and clinical data

In biomedicine, the programme includes projects that use AI to integrate data that researchers often analyse separately.

NUS and A*STAR’s MultiOmicsFM is planned as a foundation model combining DNA, RNA and gene-activity data from Singapore’s multi-ethnic datasets. Its research goals include studying disease risk and improving the design and delivery of mRNA therapies.

The BloodCounts! project is taking a different route: it will develop a DeepCBC model using the full raw parameters from complete blood-count tests together with blood-smear images. The aim is to extract more diagnostic information from a widely used test and investigate whether the combined data can help predict risks associated with conditions such as stroke and cancer.

These are prospective research applications, not a claim that a routine blood test can currently diagnose those conditions. Any clinical use would require validation, regulatory review and evidence that the model performs reliably across relevant populations and care settings.

Digital twins bring AI into agriculture and climate resilience

Another project will create knowledge-guided agricultural digital twins for Southeast Asian farmland. These virtual representations are intended to combine observed data with crop-science principles, producing forecasts and decision tools for planting, resource use and supply chains.

The emphasis on scientific knowledge is significant. A digital twin that reflects agronomic constraints can be more interpretable and easier to challenge than an opaque prediction, although its usefulness will still depend on the quality, coverage and timeliness of the underlying data.

The eight-project portfolio also includes AI-designed interfaces intended to improve energy transport. Taken together, the projects span materials, computing, genomics, health, energy, agriculture and climate resilience rather than concentrating on one commercial sector.

The programme’s operating model: a closed loop between AI and experiments

Across the portfolio, the common model is a loop:

1. AI searches or proposes hypotheses, materials, catalysts, molecular designs, code or agricultural interventions.
2. Physical or digital experiments test those proposals through automated laboratories, measurements, software analysis or field-relevant models.
3. New evidence improves the system, making subsequent predictions more grounded in observed results.
4. Domain experts decide what matters, including whether an output is manufacturable, reliable, clinically useful or environmentally relevant.

The Materials Data Foundry makes this loop most tangible because it is explicitly designed to connect AI predictions with automated synthesis and measurement. The same principle appears in catalyst screening, software verification and biomedical modelling: faster predictions are valuable only when they lead to trustworthy evidence.

Why Singapore is investing in AI4S

The initiative is also a talent and infrastructure strategy. NRF has positioned AI4S around collaboration between AI researchers and scientists in fields such as materials science, life science and quantum science. Its stated ambition is to develop “bilingual” researchers who understand both computational methods and the scientific problems to which they are applied.

The projects bring together Singaporean institutions—including NUS, NTU and A*STAR—with international partners such as the University of Toronto, Imperial College London, the University of Cambridge and Kyoto University. Imperial’s two programmes are hosted through Imperial Global Singapore at the CREATE campus, with an emphasis on international research collaboration and access.

That structure gives Singapore a way to compete for AI-driven research: fund ambitious challenge projects, build shared experimental and digital infrastructure, attract global collaborators and train researchers who can translate between models and laboratories.

What success will look like

AI4S will be judged less by the number of predictions its models produce than by what survives contact with reality. The strongest outcomes would include:

• reproducible manufacturing recipes for advanced materials;
• catalysts that perform under realistic operating conditions;
• software tools that can demonstrate meaningful correctness or reliability guarantees;
• biomedical models validated on appropriate datasets and clinical settings; and
• agricultural decision tools that improve forecasts or resource planning in practice.

Singapore’s programme is therefore placing a bet on AI connected to experiments, verification and domain expertise. The S$120 million initiative could make scientific discovery more systematic and accelerate the path from idea to prototype. But the central question remains open: whether its eight research projects can turn promising AI-assisted predictions into scalable, safe and useful technologies.