Accelerated Materials Development: Reflections from Singapore

On December 1st, I returned full-time to MIT in Boston, USA from a multi-year tour of duty in Singapore. I was hosted at the Singapore-MIT Alliance for Research & Technology (SMART), with a shorter secondment at A*STAR. Here are a few learnings worth sharing:

Learning #1.?By combining machine learning (ML) and high-throughput experiments (HTE), we achieved an R&D acceleration factor of ~20x across a wide range of materials systems — which invites us to rethink how we perform R&D.

During a leave of absence from SMART/MIT, I served as founding director of the Accelerated Materials Development for Manufacturing (AMDM) programme, a S$24.7M, 5-year research centre based at A*STAR, and including Singapore universities NUS and NTU. I enjoyed leading an amazing team to combine ML with high-throughput research methods in a diverse set of fields, from semiconductor materials and devices, thermoelectrics, polymers, nanoparticles, and more. The core hypothesis we tested, was to use data science to drive high-throughput research methods, with an emphasis on streamlining, or simplifying the research process.

In many cases, this invited a re-think of our entire R&D pipeline, from hypothesis generation to equipment design to dissemination. For example, ML+HTE allowed us to simplify our experimental apparatus to generate machine-interpretable data, affording opportunities to reduce hardware cost and complexity. Additionally, dissemination of all materials as open source (code and datasets) enables rapid reproduction of results by other groups, but also demand greater management engagement to ensure that all researchers’ careers benefit from campaign-style teamwork.

There are now several examples where ML accelerates novel materials development by 4x to 600x, with a median acceleration over 20x. The AMDM programme has been in?Kedar Hippalgaonkar's capable hands since 2019; be on the lookout for a Fall MRS video showcasing the AMDM team’s amazing work.

The use cases have progressed beyond academic hero experiments and into industrial applications. A joint A*STAR-SMART team spun off Xinterra, which applies this toolkit to accelerate the development of materials for sustainability. Xinterra is led by industry veteran and Sloan Fellow?Patrick Teyssonneyre, who previously spearheaded the global R&D initiatives of Braskem, which included the development of renewable-based chemicals.

Left: Summary of acceleration factors from several peer-reviewed research papers, including various combinations of ML and high-throughput research tools. Right: Example of accelerated nanoparticle synthesis [link to paper], which combines an active-learning algorithm (for rapid colour matching) and a regression algorithm (enabling inverse design).

Learning #2: For many researchers today, “islands of automation” may be preferable to “fully autonomous self-driving labs”

Personal take: For early-stage research, I often found “islands of automation” preferable to “fully autonomous labs” — at least for now. Fully-integrated, large autonomous labs are enticing (especially for R&D management; they make cool movies!), but have high barriers to entry for new researchers, chug consumables, and require teams of dedicated engineers to maintain and operate. The AMDM flagship lab includes a fully automated spin-coating system, with integrated polymer synthesis capability. I'm really proud of this tool and the team! But this is not an entry-level tool — it's best reserved for problems where a degree of sophistication (e.g., inert ambient) or repetition (massive parallelization of a fixed workflow) is required, resources are available to support the tool (read: several full-time staff), and the workflow is well established the old-fashioned way (read: years of experience with the traditional workflow). Over the last few years, other PIs shared stories of struggles: A postdoc spending months to get a custom-designed autonomous lab to unscrew a precursor bottlecap, of a half-year installation delay to reinforce the lab floor for the weight of an autonomous lab, a ~$33k autonomous lab control experiment… Few appreciate this.

Back in 2018, we also had the good luck/foresight to budget for smaller autonomous tools (in addition to the flagship autonomous lab). One of these was a simple liquid pipetting system. Like a Drone, it’s not as powerful or versatile as a Battleship. Initially envisioned as a training tool, it quickly became the go-to tool for the lab because of its low barrier to entry. Its utilization rate is really high, researchers are keen to jump on the tool to test things out. We also home-built new high-throughput equipment, which are usually simpler, cheaper, and faster than standard lab equipment.

Now there's a "Drone Swarm," capable of adapting to many workflows. The Drone Swarm saved us many times during COVID. When one lab temporarily shut down, another’s drones could adapt and continue where one workflow dropped off. Papers were published during the pandemic, thanks to parts of the workflow being temporarily outsourced to high-throughput equipment in other countries. I carried some of these learnings back to MIT: In Fall 2020, when I taught applied ML course and our MIT labs were down, the team conducted autonomous experiments using a Drone at SMART in Singapore, with students writing active-learning code thousands of kilometers away.

For now, the Drone Swarm is beating the fully autonomous lab in the publication race, by a respectable margin. I expect that the autonomous lab will soon produce high-impact papers out of the Drone Swarm’s reach. But there are important lessons here. If you’re a young research team that wants to ramp quickly, consider getting a Drone Swarm that leverages ML to the fullest, automating just the most time-intensive steps of your process flow, and leveraging human adaptivity in between. To read more about this “islands of automation” lab-design philosophy, see the online Supplemental Info of?this paper [direct SI link], or read excellent related work by Jason Hein’s and Curtis Berlinguette’s teams on related philosophies here and here.

The science show must go on (even during a pandemic)! One part of the “Drone Swarm” of smaller, geographically-distributed semi-autonomous research tools, enabling resilient and adaptive workflows. This SMART-based OpenTrons tool was used for teaching a remote MIT course in Fall 2020.

Learning #3: Upskilling works, but we need dedicated training materials to create hybrid materials/chem researchers + data scientists.

Third, I saw many motivated materials researchers become ML proficient with proper training and community support. At the very start of AMDM, we were told we could not make external hires, but needed to retrain staff already at A*STAR. So?Kedar?and I, together with ML experts on the team, developed an ML training curriculum for materials researchers, combining online sources, seminars, code repositories, workshops, and hackathons. We made a point to showcase early success stories at local workshops, at international conference like the MRS, sharing datasets and code, so other researchers were able to more quickly apply ML to their own research problems. We created new courses at NTU and MIT. And the community grew. An estimated 1% of Singapore’s R&D workforce attended one or more workshop, upskilling and creating professional connections that led to future research projects, and a vibrant AccMatDev biweekly online seminar series and community. One early hackathon winner even joined our team, later being promoted to group leader. Through this process, ML experts working with several materials research teams applied inductive reasoning to identify trends and improve customization of algorithms to meet materials researchers needs. It really invites imagining, what multi-year challenges we could create for materials and chemistry, to mimic the success of the biennial protein-structure prediction challenge (CASP), the competition that helped inspire AlphaFold v2.0.

Left: YouTube video channel, disseminating success stories with associated datasets and code. Right: 2018 Applied ML?workshop & hackathon, disseminating code and datasets from recent success cases applying ML to materials research. The hackathon winner eventually joined our AMDM team and became a team leader. An estimated 1% of Singapore’s R&D workforce attended one or more such workshop.

Learning #4: International engagement is key to addressing global challenges

To create AMDM, international engagement was key. The merging of ideas and perspectives led to the original AMDM leadership team of myself (SMART),?Kedar?(A*STAR), and?Xiaonan?(then NUS; now Tsinghua), together with support of visionary researchers, reviewers, mentors, and administrators along the way, enabled us to build an international Community of researchers to challenge the way research is done. Once several successful cases emerged, the core AMDM philosophy became accepted and incorporated into several other research programmes around Singapore and around the world. But at the beginning, we faced considerable skepticism. Could ML be adapted to smaller datasets? Could high-throughput experiments also produce high-quality materials? The AMDM Community overcame these obstacles by engaging a diverse range of international perspectives.

While some pandemic-era rhetoric suggests that such diverse Community-building can be achieved in a vacuum, that’s not the reality I observed first-hand. It was an international SMART-A*STAR collaboration that led to AMDM. An even broader international collaboration led by?Alán Aspuru-Guzik?resulted in the Acceleration Consortium out of UofT. Earlier still, an international collaboration spearheaded by?Hermann Tribukait led to Mission Innovation's IC6. I strongly believe that international engagement will create new Communities to address future challenges. To create new Communities like AMDM, a posture of appropriate international engagement is essential. Diverse ideas do not spontaneously appear inside of an echo chamber, and global problems require global solutions.

Lowering the barrier to entry for newcomers by creating an inclusive environment with available code and datasets is a start, but we can go further. We can engage international colleagues to write perspective pieces and review articles, involving students in this learning process. Once travel is again available, we can encourage our students to travel abroad and build bridges with peers around the world. We can strive to cover more geographically diverse articles in our journal clubs, and include geographically diverse references in our next research articles (thanks?Steve Cranford?for inspiring this ideas with the recent Matter DEI checklist).

In closing, I’m very grateful for these opportunities to serve, to lead, and to learn. I wish my Singapore- and Asia-based colleagues and friends all the best, as I return my full focus to MIT starting December 2021. Safe onward journey to all, and hoping for many fruitful future collaborations!