1*e28069aeSJunchao Zhang--- 2*e28069aeSJunchao Zhangorphan: true 3*e28069aeSJunchao Zhang--- 4*e28069aeSJunchao Zhang 5*e28069aeSJunchao Zhang(2026-feb-cass-petsc-bof)= 6*e28069aeSJunchao Zhang 7*e28069aeSJunchao Zhang*2026 Consortium for the Advancement of Scientific Software ([CASS](https://cass.community/)) BoF Days -- the PETSc/TAO Session (Virtual via Zoom)* 8*e28069aeSJunchao Zhang 9*e28069aeSJunchao Zhang 10*e28069aeSJunchao Zhang# PETSc/TAO: Recent Advances, User Experiences, and Community Discussion 11*e28069aeSJunchao Zhang 12*e28069aeSJunchao Zhang**Time:** Feb 11, 2026, 10:00am - 11:30am (Central Time, US and Canada) 13*e28069aeSJunchao Zhang 14*e28069aeSJunchao Zhang**Registration:** [https://argonne.zoomgov.com/meeting/register/ay4bMcRgSZaZ-l7u9AzAzQ](https://argonne.zoomgov.com/meeting/register/ay4bMcRgSZaZ-l7u9AzAzQ) 15*e28069aeSJunchao Zhang 16*e28069aeSJunchao Zhang**Note:** All participants **must register** to receive their own Zoom link. The step is brief and requires only a name and email address. The meeting will not be recorded. 17*e28069aeSJunchao Zhang 18*e28069aeSJunchao Zhang## Agenda (tentative): 19*e28069aeSJunchao Zhang 20*e28069aeSJunchao Zhang| Time | Speaker | Affiliation | Title | 21*e28069aeSJunchao Zhang| :---- | :---- | :---- | :---- | 22*e28069aeSJunchao Zhang| 10:00 AM | Barry Smith | Argonne National Laboratory | Welcome and introduction | 23*e28069aeSJunchao Zhang| 10:05 AM | Vincent Robert | CEA Paris-Saclay and Sorbonne Université | [Solving saddle-point systems from contact mechanics in HPC context using PETSc](#talk-vincent-robert) | 24*e28069aeSJunchao Zhang| 10:12 AM | Bahaâ Sidi | EDF R\&D, ENPC, and Sorbonne Université | [Scalable domain decomposition solvers for Hybrid High-Order methods in PETSc](#talk-bahaa-sidi) | 25*e28069aeSJunchao Zhang| 10:19 AM | Barry Smith, Richard Tran Mills, Hansol Suh, Junchao Zhang | Argonne National Laboratory | [PETSc new feature update](#talk-petsc-anl) | 26*e28069aeSJunchao Zhang| 10:42 AM | Xiaodong Liu | Remcom Inc | [A PETSc-based vector finite element method code for solving the frequency-domain Maxwell equations](#talk-xiaodong-liu) | 27*e28069aeSJunchao Zhang| 10:49 AM | Jeremy L Thompson | University of Colorado Boulder | [Ratel \- Using MI300A APUs with libCEED and PETSc](#talk-jeremy-thompson) | 28*e28069aeSJunchao Zhang| 10:56 AM | Darsh Nathawani | Louisiana State University | [PETSc in Proteus](#talk-darsh-nathawani) | 29*e28069aeSJunchao Zhang| 11:03 AM | All participants | | Users take this chance to give feedbacks and make feature requests | 30*e28069aeSJunchao Zhang| 11:08 AM | Barry Smith | Argonne National Laboratory | Open the PETSc/AI discussion | 31*e28069aeSJunchao Zhang| 11:10 AM | Mark Adams | Lawrence Berkeley National Laboratory | [Using AI code assistants for large projects with PETSc](#talk-mark-adams) | 32*e28069aeSJunchao Zhang| 11:17 AM | All participants | | Open discussion about PETSc in the era of AI and large language models | 33*e28069aeSJunchao Zhang| 11:30 AM | BoF concludes | | | 34*e28069aeSJunchao Zhang 35*e28069aeSJunchao Zhang 36*e28069aeSJunchao Zhang 37*e28069aeSJunchao Zhang## List of Abstracts: 38*e28069aeSJunchao Zhang 39*e28069aeSJunchao Zhang(talk-vincent-robert)= 40*e28069aeSJunchao Zhang 41*e28069aeSJunchao Zhang:::{topic} **Solving saddle-point systems from contact mechanics in HPC context using PETSc** 42*e28069aeSJunchao Zhang 43*e28069aeSJunchao Zhang**Speaker:** Vincent Robert 44*e28069aeSJunchao Zhang 45*e28069aeSJunchao Zhang**Affiliation:** CEA Paris-Saclay and Sorbonne Université 46*e28069aeSJunchao Zhang 47*e28069aeSJunchao Zhang**Abstract:** The French Alternative Energies and Atomic Energy Commission (CEA), is one of the leading French public research organisations. MANTA is a new CEA open-source industrial-strength simulation code which is dedicated to the solving, in a high performance computing context, of partial differential equations mainly from mechanics of deformable structures and their interactions. MANTA relies on PETSc to solve all its linear systems. Using PETSc, we focus on developing numerical methods to solve saddle-point systems arising from large-scale 3D contact problems, with the goal to include them in MANTA. Our approach is based on a projection method which transforms the indefinite saddle-point system into a positive definite reduced system. We propose several preconditioning matrices, introduce a scaling factor that improves the conditioning of the preconditioned system, and fix several issues related to the parallel distribution of the matrices. Using the Conjugate Gradient method as the iterative linear solver with an Algebraic Multigrid Method for the preconditioning, the current numerical tests show great scalability potential of the method. 48*e28069aeSJunchao Zhang::: 49*e28069aeSJunchao Zhang 50*e28069aeSJunchao Zhang 51*e28069aeSJunchao Zhang(talk-bahaa-sidi)= 52*e28069aeSJunchao Zhang 53*e28069aeSJunchao Zhang:::{topic} **Scalable domain decomposition solvers for Hybrid High-Order methods in PETSc** 54*e28069aeSJunchao Zhang 55*e28069aeSJunchao Zhang**Speaker:** Bahaâ Sidi 56*e28069aeSJunchao Zhang 57*e28069aeSJunchao Zhang**Affiliation:** EDF R\&D, ENPC, and Sorbonne Université 58*e28069aeSJunchao Zhang 59*e28069aeSJunchao Zhang**Abstract:** This talk presents ongoing work carried out at EDF R&D on the coupling of Hybrid High-Order (HHO) discretizations with scalable linear solvers available in PETSc. The objective is to efficiently solve large-scale linear elasticity problems arising from HHO methods, which naturally lead to block-structured systems separating cell and face unknowns. We focus on Schur complement strategies based on PCFIELDSPLIT, where the cell block is eliminated locally and the resulting interface problem is solved using a combination of explicit Schur complements, Krylov methods, and domain decomposition preconditioners. 60*e28069aeSJunchao Zhang::: 61*e28069aeSJunchao Zhang 62*e28069aeSJunchao Zhang 63*e28069aeSJunchao Zhang(talk-petsc-anl)= 64*e28069aeSJunchao Zhang 65*e28069aeSJunchao Zhang:::{topic} **PETSc new feature update** 66*e28069aeSJunchao Zhang 67*e28069aeSJunchao Zhang**Speaker:** Barry Smith, Richard Tran Mills, Hansol Suh, Junchao Zhang 68*e28069aeSJunchao Zhang 69*e28069aeSJunchao Zhang**Affiliation:** Argonne National Laboratory 70*e28069aeSJunchao Zhang 71*e28069aeSJunchao Zhang**Abstract:** PETSc developers at Argonne will give an update on new features in PETSc, including the new PETSc Fortran bindings, PetscRegressor, TaoTerm, PETSc/GPU backends, mixed-precision and out-of-core facility support in PETSc/MUMPS, and integration with OpenFOAM via petsc4foam, among others. 72*e28069aeSJunchao Zhang::: 73*e28069aeSJunchao Zhang 74*e28069aeSJunchao Zhang 75*e28069aeSJunchao Zhang(talk-xiaodong-liu)= 76*e28069aeSJunchao Zhang 77*e28069aeSJunchao Zhang:::{topic} **A PETSc-based vector finite element method code for solving the frequency-domain Maxwell equations** 78*e28069aeSJunchao Zhang 79*e28069aeSJunchao Zhang**Speaker:** Xiaodong Liu 80*e28069aeSJunchao Zhang 81*e28069aeSJunchao Zhang**Affiliation:** Remcom Inc 82*e28069aeSJunchao Zhang 83*e28069aeSJunchao Zhang**Abstract:** A PETSc-based vector finite element method code for solving the frequency-domain Maxwell equations 84*e28069aeSJunchao ZhangAbstract: A PETSc-based vector finite element method (FEM) code is under active development at Remcom Inc. for solving the frequency-domain Maxwell equations. The current implementation supports key electromagnetic analysis capabilities, including lumped-port excitation, perfectly matched layer (PML) boundary conditions, far-field radiation pattern computation, and input impedance evaluation. 85*e28069aeSJunchao Zhang 86*e28069aeSJunchao ZhangThe code is built on PETSc’s DMPlex infrastructure to manage fully unstructured meshes, enabling scalable assembly and solution workflows. Direct solvers based on MUMPS have been validated on problems with up to 13 million complex-valued degrees of freedom, with scalability currently constrained by available computational resources. 87*e28069aeSJunchao Zhang 88*e28069aeSJunchao ZhangOngoing work is twofold. First, we are investigating iterative solvers, in particular BDDC-preconditioned FGMRES, where additional challenges arise from the vector-valued nature of the FEM discretization. Second, future development will target expanded electromagnetic modeling capabilities and improved solver robustness and scalability, with an emphasis on industrial applicability and alignment with emerging exascale computing platforms. 89*e28069aeSJunchao Zhang::: 90*e28069aeSJunchao Zhang 91*e28069aeSJunchao Zhang 92*e28069aeSJunchao Zhang(talk-jeremy-thompson)= 93*e28069aeSJunchao Zhang 94*e28069aeSJunchao Zhang:::{topic} **Ratel - Using MI300A APUs with libCEED and PETSc** 95*e28069aeSJunchao Zhang 96*e28069aeSJunchao Zhang**Speaker:** Jeremy L Thompson 97*e28069aeSJunchao Zhang 98*e28069aeSJunchao Zhang**Affiliation:** University of Colorado Boulder 99*e28069aeSJunchao Zhang 100*e28069aeSJunchao Zhang**Abstract:** AMD's MI300A APUs have unified memory accessible by both the CPU and APU cores on the accelerator, eliminating the need for communication between device and host memory spaces. In this talk, we discuss our work with the solid mechanics library Ratel to use this unified memory model for FEM and material point method (MPM) problems. Ratel uses PETSc's mesh management and solvers while relying upon libCEED for the local action of the linear and non-linear operators in a matrix-free fashion. Ratel has successfully scaled simulations to tens of billions of material points with the MI300As. 101*e28069aeSJunchao Zhang::: 102*e28069aeSJunchao Zhang 103*e28069aeSJunchao Zhang 104*e28069aeSJunchao Zhang(talk-darsh-nathawani)= 105*e28069aeSJunchao Zhang 106*e28069aeSJunchao Zhang:::{topic} **PETSc in Proteus** 107*e28069aeSJunchao Zhang 108*e28069aeSJunchao Zhang**Speaker:** Darsh Nathawani 109*e28069aeSJunchao Zhang 110*e28069aeSJunchao Zhang**Affiliation:** Louisiana State University 111*e28069aeSJunchao Zhang 112*e28069aeSJunchao Zhang**Abstract:** Proteus is a mixed language toolkit that offers models for continuum mechanical processes described by PDEs. It provides CG and DG discretization, classic linear and nonlinear solvers, and time integrators. PETSc is primarily integrated to use Vec and Mat objects as well as KSP and PC to build solvers. Currently, we are expanding PETSc support by using DMPlex for our meshing framework. 113*e28069aeSJunchao Zhang::: 114*e28069aeSJunchao Zhang 115*e28069aeSJunchao Zhang 116*e28069aeSJunchao Zhang(talk-mark-adams)= 117*e28069aeSJunchao Zhang 118*e28069aeSJunchao Zhang:::{topic} **Using AI code assistants for large projects with PETSc** 119*e28069aeSJunchao Zhang 120*e28069aeSJunchao Zhang**Speaker:** Mark Adams 121*e28069aeSJunchao Zhang 122*e28069aeSJunchao Zhang**Affiliation:** Lawrence Berkeley National Laboratory 123*e28069aeSJunchao Zhang 124*e28069aeSJunchao Zhang**Abstract:** I will discuss my experience over the past 4 months using AI code assistants (eg, Claude, Gemini, GPT5), integrated with IDEs (eg, VS Code with Roo) for large projects using or developing in PETSc. I have used chatGPT for several months for data processing and very small coding projects. GPT does not seem as good as Claude and Gemini at scientific code, such as PETSc, but I still find it useful for very quick problems. 125*e28069aeSJunchao Zhang 126*e28069aeSJunchao Zhang 127*e28069aeSJunchao ZhangI will give a quick demonstration and discuss my experience developing a data assimilation tool in PETSc for shallow water equations—both new topics for me. I will provide tips on using these tools and give an idea of how well they currently support developing PETSc C codes. 128*e28069aeSJunchao Zhang::: 129