About
I am a final-year Ph.D. candidate in Computer Science at the University of Minnesota, advised by Yousef Saad in the Scientific Computing Lab.
My work sits where numerical linear algebra meets machine learning: learned preconditioners and algebraic multigrid for sparse linear systems, graph coarsening, and, more recently, RL and LLM agents whose rewards come from real numerical solves. I spent two summers at the Center for Applied Scientific Computing at Lawrence Livermore National Laboratory.
Research
Learned multilevel preconditioners
Graph neural networks that learn smoothers and grid-transfer operators on top of an AMG hierarchy, used as drop-in preconditioners for Krylov solvers on nonsymmetric and indefinite systems.
RL and LLM agents for solver design
Coarse/fine splitting as a sequential decision problem: from graph RL (CoarseRL), to an LLM that steers RL training under a frozen solver-based evaluation, with GRPO post-training of the LLM on these verifiable rewards as the next step.
Graph coarsening
Connecting multilevel ideas from scientific computing with graph learning: spectral coarsening that speeds up graph classification, and multilevel training of GNNs.
Low-precision eigensolvers
Orthogonalization-free Rayleigh–Ritz projections that stay accurate when the basis is built in single or half precision.
Publications
-
KDD 2026Graph Neural Multilevel Preconditioners for Iterative SolversZechen Zhang, Rui Peng Li, Yousef SaadGMP reuses an AMG hierarchy as a structural prior and learns relaxation with message-passing smoothers and grid transfer with bipartite cross-attention. Across 867 SuiteSparse non-SPD matrices it wins 63.5% of head-to-head comparisons against a single-level learned preconditioner.
-
SIAM J. Sci. Comput. 2026Mixed Precision Orthogonalization-Free Projection Methods for Eigenvalue and Singular Value ProblemsTianshi Xu, Zechen Zhang, Jie Chen, Yousef Saad, Yuanzhe XiOFRR extracts accurate eigen- and singular-value approximations from non-orthogonal bases built in reduced precision; with a Hessenberg-process basis it keeps errors below 10−4 in all-FP16 arithmetic.
-
ICLR 2026 WorkshopCoarseRL: A Graph Reinforcement Learning Method for Algebraic Multigrid CoarseningSoha Yusuf, Zechen Zhang, Kowshik Thopalli, Rui Peng LiAMG coarse/fine splitting as sequential node selection on the matrix graph, with a diagonal-dominance constraint on the fine–fine block. Workshop on AI and PDEs.
-
SeMA Journal 2022
-
In progressAgentic AMG Coarsening with Verifiable-Reward RLA graph RL agent selects coarse points and is rewarded by the convergence and cost of the resulting two-grid method. An LLM in the outer loop proposes how to retrain it, and a frozen, solver-based evaluation keeps only Pareto-optimal agents. Next: GRPO post-training of the LLM on the same verifiable solver rewards.
Experience & Education
CoarseRL: graph reinforcement learning for AMG coarsening. Workshop paper at ICLR 2026 (AI and PDEs).
MT-GNN: multilevel (FAS) training of graph neural networks via graph coarsening.