FigureAsia 35 Under 35 · Science
Priya Donti
Age 33 · Machine learning for power grids · India / United States
Climate-AI researcher building realistic benchmarks and decision tools for low-carbon power systems.
- Approximate age at the edition eligibility date
- 33
- Field
- Climate and energy systems
- Country or region
- India / United States
- FigureAsia U35 Assessment
- 90.8 / 100
Profile
Career and documented record
Priya Donti works on the operational layer of decarbonisation: how power systems absorb more renewable generation without sacrificing reliability. In 2025, as senior author, she helped introduce RL2Grid, a reinforcement-learning benchmark developed with French grid expertise to represent the constraints, safety boundaries and operating heuristics that academic simulations often omit.
The contribution is infrastructure rather than spectacle. By unifying realistic environments, baselines and evaluation, RL2Grid makes it harder for an algorithm to appear successful only because the test was forgiving. It also creates a shared path for researchers to compare learning-based control against established grid practice.
Donti's co-founding of Climate Change AI and her wider work on optimisation-informed learning give the project institutional reach. The science remains bounded: a benchmark is not permission to operate a live grid. Its value is in making the evidence required for that transition more demanding.
FigureAsia selection
Why Priya Donti is on the list
Donti combines methodological depth with one of climate technology's hardest deployment contexts. RL2Grid earns credit for bringing safety, physical constraints and incumbent practice into the evaluation itself—exactly where much applied AI is weakest.
Verified work
The 2025–26 record
RL2Grid
Senior-authored a realistic reinforcement-learning benchmark for power-grid operations.
Operational grounding
Integrated constraints, safety boundaries, heuristics and common baselines with grid-domain collaborators.
Climate research infrastructure
Extended an open programme connecting machine learning with deployable decarbonisation tools.
Field context
The work in its field
Electric grids are constrained dynamical systems in which an attractive average result can still hide an unacceptable failure. Useful AI must therefore be evaluated against physics, contingencies and operator practice.
FigureAsia U35 Assessment
Assessment breakdown
90.8out of 100
Substantive 2025–2026 contribution
17.4 / 20
Senior-authored a realistic reinforcement-learning benchmark for power-grid operations.
Verified scientific impact
13.4 / 15
RL2Grid supplies shared infrastructure for a high-consequence field and is grounded in real operator constraints.
Originality and distinction
9.1 / 10
The distinction lies in embedding grid safety and operational heuristics into a unified learning benchmark instead of evaluating control in idealised isolation.
Field influence
9.1 / 10
Researchers in machine learning for power grids now have a stronger result to test, extend or challenge because of this contribution.
Individual agency
9.3 / 10
Donti is the senior author and leads the academic programme, while utility and co-author contributions remain explicit.
Durability and trajectory
4.7 / 5
The contribution builds on an active line of work at MIT EECS and LIDS, with further tests and applications still to come.
Asian significance and global relevance
4.7 / 5
Indian-American scientist who spent part of her childhood in India and now leads climate-and-AI research at MIT.
Evidential validity and reproducibility
7.3 / 8
The benchmark exposes physical constraints and comparable baselines; no claim of live deployment is made.
Advance in scientific knowledge
6.5 / 7
It clarifies which apparent gains survive closer approximation to real power-system operations.
Translational or methodological utility
4.6 / 5
Researchers and utilities gain a common test bed for safer, more comparable grid-control work.
Responsible research stewardship
4.7 / 5
Safety boundaries and incumbent operator knowledge are treated as design inputs rather than obstacles to optimisation.