In this case study, the National University of Singapore’s I-FIM (Singapore’s 6th Research Centre of Excellence) and the University of Toronto’s Acceleration Consortium partnered to launch the Materials Data Foundry (MDF), an open autonomous lab combining multi-modal AI, high-throughput robotics, and in-situ measurements. Funded under Singapore’s S$120M AI-for-Science Initiative with industrial partners NVIDIA and VeChain, the MDF aims to build the world’s largest experimental materials property dataset — 50,000 high-fidelity experiments. We evaluated this partnership to help your team structure similarly ambitious open-science infrastructure collaborations.
1. Executive Summary
In June 2026, NUS I-FIM and the University of Toronto Acceleration Consortium launched the Materials Data Foundry, an open autonomous laboratory designed to address the primary barrier to AI-driven materials discovery: the lack of high-quality, standardised experimental data. The S$10M project is one of eight inaugural initiatives under Singapore’s S$120M AI-for-Science Initiative and brings together complementary expertise — NUS in functional materials, U of T in self-driving laboratories — with industrial partners NVIDIA contributing compute infrastructure and VeChain providing blockchain-based data integrity.
- Subject: NUS I-FIM (Singapore’s 6th Research Centre of Excellence) and U of T Acceleration Consortium
- Problem: Lack of high-quality standardised materials data is the primary barrier to AI-driven materials discovery
- Solution: Open autonomous lab building the world’s largest experimental materials property dataset — 50,000 high-fidelity experiments
- Result: Open dataset, multi-modal foundation models, APIs, and protocols for scalable automated labs
2. The Challenge
AI-driven materials discovery holds enormous potential but faces a fundamental data bottleneck: machine learning models require massive, standardised datasets that capture both synthesis recipes and measured outcomes — but no single laboratory can produce enough high-quality data to train general-purpose materials foundation models. Compounding this, materials research suffers from a reproducibility crisis, with synthesis protocols and characterisation data poorly captured and inconsistently formatted across institutions.
- Data scarcity at scale: AI models need massive, standardised datasets that no single lab can produce alone — existing materials databases are too small, noisy, and inconsistent for modern machine learning
- Reproducibility problem: Materials research lacks standardised data capture — synthesis recipes, processing parameters, and measured outcomes are rarely captured together in machine-readable formats
- Cross-continent coordination: Singapore-Toronto operations add logistical complexity that single-site labs avoid, requiring integrated instrument control, unified data schemas, and aligned research protocols across 12 time zones
Both sides recognised they needed each other: NUS for U of T’s proven self-driving laboratory platform and autonomous experimentation methodology, U of T for NUS’s world-class functional materials expertise and Singapore’s strategic investment in AI-for-Science infrastructure.
3. The Strategy
Rather than building a conventional collaborative research programme with separate workstreams, NUS and U of T designed the Materials Data Foundry as a single open autonomous laboratory combining multi-modal AI with high-throughput robotics to generate data at industrial scale — with all protocols, datasets, and APIs released openly for replication by other laboratories worldwide.
- Open autonomous lab architecture: Multi-modal AI integrated with high-throughput robotics and in-situ measurements to generate standardised data at industrial scale — each experiment captures synthesis recipe, processing parameters, and measured properties in a unified schema
- Cross-continent complementary expertise: NUS contributes functional materials research and Southeast Asia’s materials science ecosystem; U of T contributes self-driving laboratory methodology and North American autonomous lab infrastructure
- Industrial partner infrastructure: NVIDIA provides compute platform for AI model training and simulation; VeChain provides blockchain data integrity layer ensuring dataset provenance and tamper-evident records for every experiment
Resources were split by comparative strength: NUS contributed functional materials expertise, laboratory facilities, and Singapore-based researchers; U of T contributed self-driving laboratory methodology, automated experimentation platforms, and Toronto-based research infrastructure. Industrial partners contributed infrastructure in kind — NVIDIA’s compute platform and VeChain’s blockchain integrity layer — rather than direct research funding.
4. The Results
The Materials Data Foundry achieved early institutional validation and established a clear trajectory toward its ambitious data generation targets, with open infrastructure designed to maximise global impact.
- Inaugural AI4S project: Selected as one of eight founding projects under Singapore’s S$120M AI-for-Science Initiative, providing government anchor funding and strategic visibility
- 50,000-experiment target: Ambitious data generation goal linking synthesis recipes to measured outcomes at a scale no single materials lab has achieved — designed to train the next generation of materials foundation models
- Open protocols and APIs: All laboratory protocols, data schemas, and application programming interfaces designed for replication by other laboratories worldwide — maximising scientific impact beyond the founding partners
The open infrastructure design is particularly significant: because the MDF’s protocols, schemas, and APIs are openly published, any research group with automated laboratory equipment can contribute compatible data, creating a network effect that compounds the value of the 50,000-experiment core dataset.
5. The Melan Approach
Melan advises structuring partnerships like this one when the primary output is open infrastructure rather than proprietary advantage — the consortium model works best when all partners benefit from a shared resource that none could build alone, and when industrial contributors participate through infrastructure rather than direct funding.
- Governance model: Joint leadership from I-FIM and Acceleration Consortium directors. Melan would add a formal industrial advisory board with decision rights over the research roadmap — ensuring industrial partners have meaningful input despite contributing infrastructure rather than funding.
- Risk allocation: Government grant funding shifts financial risk to the public sector. Melan recommends allocating budget for a dedicated commercial translation pathway — open infrastructure projects often struggle to transition discoveries from published datasets to market applications.
- Shared goal: Build open infrastructure that accelerates AI-driven materials discovery globally while establishing Singapore and Canada as leaders in autonomous laboratory research. Melan would add formal contribution metrics for industrial partners to sustain engagement beyond the initial grant period.
This open-infrastructure consortium model is replicable for other domains facing similar data bottlenecks — any scientific field where AI progress is limited by the lack of standardised, high-quality experimental datasets can benefit from a similar architecture of autonomous labs, open protocols, and cross-continent complementary expertise.
Building an open-science AI infrastructure partnership?
Melan helps research institutions and consortia structure open-infrastructure collaborations with industrial advisory boards, commercial translation pathways, and governance models that sustain engagement beyond initial grant periods.
