{"id":1158675,"date":"2026-01-06T09:43:03","date_gmt":"2026-01-06T17:43:03","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-project&#038;p=1158675"},"modified":"2026-05-14T15:39:22","modified_gmt":"2026-05-14T22:39:22","slug":"gridfm","status":"publish","type":"msr-project","link":"https:\/\/www.microsoft.com\/en-us\/research\/project\/gridfm\/","title":{"rendered":"GridFM"},"content":{"rendered":"<section class=\"mb-3 moray-highlight\">\n\t<div class=\"card-img-overlay mx-lg-0\">\n\t\t<div class=\"card-background  has-background- card-background--full-bleed\">\n\t\t\t<img loading=\"lazy\" decoding=\"async\" width=\"1536\" height=\"1024\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Designer.png\" class=\"attachment-full size-full\" alt=\"AI backkground giving a sense of power grids and foundtaional models\" style=\"object-position: 66% 43%\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Designer.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Designer-300x200.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Designer-1024x683.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Designer-768x512.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Designer-240x160.png 240w\" sizes=\"auto, (max-width: 1536px) 100vw, 1536px\" \/>\t\t<\/div>\n\t\t<!-- Foreground -->\n\t\t<div class=\"card-foreground d-flex mt-md-n5 my-lg-5 px-g px-lg-0\">\n\t\t\t<!-- Container -->\n\t\t\t<div class=\"container d-flex mt-md-n5 my-lg-5 \">\n\t\t\t\t<!-- Card wrapper -->\n\t\t\t\t<div class=\"w-100 w-lg-col-5\">\n\t\t\t\t\t<!-- Card -->\n\t\t\t\t\t<div class=\"card material-md-card py-5 px-md-5\">\n\t\t\t\t\t\t<div class=\"card-body \">\n\t\t\t\t\t\t\t\n\t\t\t\t\t\t\t\n\n<h1 class=\"wp-block-heading\" id=\"gridfm\">GridFM<\/h1>\n\n\n\n<p>Small foundation models for the electric grid<\/p>\n\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\t<\/div>\n<\/section>\n\n\n\n\n\n<p>GridFM is a Microsoft Research initiative to build <strong>small foundation models for the electric power grid<\/strong>, applying the same recipe that drives modern language and weather models to the physics of AC power flow. The first release in the family, <strong>GridSFM<\/strong>, predicts complete AC Optimal Power Flow (AC-OPF) solutions in milliseconds \u2014 bus voltages, generator dispatch, branch flows, and a feasibility verdict \u2014 directly from a grid&#8217;s topology and operating conditions.<\/p>\n\n\n\n<p>Traditional AC-OPF solvers are accurate but slow, taking minutes to hours on real-world grids with tens of thousands of components. As power systems grow more volatile under datacenter expansion, renewable variability, electrification, and extreme weather, operators need to evaluate <strong>thousands of scenarios in seconds<\/strong>, not hours. GridFM is designed to close that gap without sacrificing physical fidelity.<\/p>\n\n\n\n<p><strong>What&#8217;s new (May 2026):<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>GridSFM is open source.<\/strong> The model architecture, training pipeline, and warm-start integration with PowerModels.jl are available on <a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/github.com\/microsoft\/GridSFM\">GitHub<span class=\"sr-only\"> (opens in new tab)<\/span><\/a>; checkpoints are on <a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/huggingface.co\/collections\/microsoft\/gridsfm\">Hugging Face<span class=\"sr-only\"> (opens in new tab)<\/span><\/a>.<\/li>\n\n\n\n<li><strong>Two model tiers.<\/strong> GridSFM-Open (~15M parameters, grids up to ~4,000 buses, MIT licensed) for research and prototyping, and GridSFM-Premier (~100M parameters, grids up to ~80,000 buses) for production-scale networks. Both share the same architecture.<\/li>\n\n\n\n<li><strong>A companion open-data pipeline.<\/strong> We also released a five-stage pipeline that builds OPF-solvable transmission models for all 48 contiguous US states and six multi-state regions \u2014 including the full Western (5,076 buses) and Eastern (21,697 buses) Interconnections \u2014 using only public data (OpenStreetMap, EIA, US Census).<\/li>\n\n\n\n<li><strong>54 reproducible grid models<\/strong> are publicly released alongside the code, lowering the barrier to transmission-level research without proprietary or critical-infrastructure-restricted data.<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"150\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-1024x150.png\" alt=\"GridFM is built around four core tenets: topology-agnostic, feasibility-aware, physics-grounded, and data-efficient.\" class=\"wp-image-1172114\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-1024x150.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-300x44.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-768x112.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-1536x225.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-2048x299.png 2048w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-240x35.png 240w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">Topology Agnostic, Feasibility Aware, Physics Grounded, Data Efficient<\/figcaption><\/figure>\n\n\n\n<div class=\"wp-block-media-text has-video  has-vertical-margin-small  has-vertical-padding-none  has-media-on-the-right is-stacked-on-mobile is-vertically-aligned-top\"><div class=\"wp-block-media-text__content\">\n\n\n\n<p><em>Read the white papers:<\/em> <a href=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/05\/GridFM_white_paper.pdf\">GridSFM white paper (PDF)<\/a> \u00b7 <a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2605.04289\">Open-data pipeline (arXiv 2605.04289)<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n\n\n<h2 class=\"wp-block-heading\" id=\"motivation-behind-the-research\">Motivation behind the research<\/h2>\n\n\n\n<div class=\"wp-block-media-text has-vertical-margin-small  has-vertical-padding-none  is-stacked-on-mobile is-vertically-aligned-top\"><figure class=\"wp-block-media-text__media\"><img loading=\"lazy\" decoding=\"async\" width=\"904\" height=\"581\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Screenshot-2025-12-16-110540.png\" alt=\"example model of Texas power grid (ERCOT)\" class=\"wp-image-1158695 size-full\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Screenshot-2025-12-16-110540.png 904w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Screenshot-2025-12-16-110540-300x193.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Screenshot-2025-12-16-110540-768x494.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/Screenshot-2025-12-16-110540-240x154.png 240w\" sizes=\"auto, (max-width: 904px) 100vw, 904px\" \/><\/figure><div class=\"wp-block-media-text__content\">\n<p>The foundation of power-system operations is an optimization problem: determining the generator dispatch that minimizes cost while satisfying thousands of physical and operational constraints. This AC Optimal Power Flow problem must be solved <strong>every 5\u201315 minutes for real-time dispatch, hourly and daily for electricity markets, and across thousands of contingencies for security assessment<\/strong>.<\/p>\n\n\n\n<p>The AC power-flow equations are non-convex and nonlinear. Interior-point solvers like IPOPT handle them reliably, but require minutes to hours per solve on large grids \u2014 especially when uncertainty in load and generation has to be accounted for. Solving a single AC-OPF for one operating point is tractable. <strong>Solving it across thousands of scenarios for planning, contingency screening, or market clearing is not.<\/strong><\/p>\n\n\n\n<p>Grid complexity is increasing on every axis:<\/p>\n\n\n\n<p><strong>Extreme-weather contingencies<\/strong> demand fast scenario sweeps.<\/p>\n\n\n\n<p><strong>Renewable variability<\/strong> introduces frequent re-dispatch.<\/p>\n\n\n\n<p><strong>Distributed energy resources<\/strong> add decision variables.<\/p>\n\n\n\n<p><strong>Datacenter load growth<\/strong> and <strong>electrification<\/strong> stress the network in new spatial patterns.<\/p>\n<\/div><\/div>\n\n\n\n<div class=\"wp-block-media-text has-vertical-margin-small  has-vertical-padding-none  has-media-on-the-right is-stacked-on-mobile is-vertically-aligned-top\" style=\"grid-template-columns:auto 40%\"><div class=\"wp-block-media-text__content\">\n<p>The computational cost of conventional AC-OPF is becoming a limiting factor for operators, planners, and researchers alike. GridFM aims to make this loop fast enough to be interactive: a single forward pass of a neural surrogate, evaluated in milliseconds, with feasibility classification built in so planners can route only the borderline scenarios to a full solver.<\/p>\n\n\n\n<p>A second, structural problem holds back research and education on the US grid: <strong>detailed transmission network data in the United States is restricted under critical-infrastructure regulations<\/strong>. Academic and policy researchers historically depend on a small set of fictitious or anonymized test cases. GridFM&#8217;s open-data pipeline addresses this directly by constructing complete, OPF-solvable models entirely from public sources, so that students, policymakers, and researchers can study realistic grid behavior without proprietary data.<\/p>\n<\/div><figure class=\"wp-block-media-text__media\"><img loading=\"lazy\" decoding=\"async\" width=\"924\" height=\"1024\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/GridFM-Topology-edited-924x1024.png\" alt=\"GridFM Topology diagram\" class=\"wp-image-1158697 size-full\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/GridFM-Topology-edited-924x1024.png 924w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/GridFM-Topology-edited-271x300.png 271w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/GridFM-Topology-edited-768x851.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/GridFM-Topology-edited-1386x1536.png 1386w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2025\/12\/GridFM-Topology-edited-162x180.png 162w\" sizes=\"auto, (max-width: 924px) 100vw, 924px\" \/><\/figure><\/div>\n\n\n\n<p><\/p>\n\n\n\n<p><\/p>\n\n\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-9d6595d7 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\" style=\"flex-basis:66.66%\">\n<h3 class=\"wp-block-heading\" id=\"gridsfm-is-designed-around-four-core-tenets\">GridSFM is designed around four core tenets:<\/h3>\n\n\n\n<p><strong>Topology Agnostic.<\/strong> A single model with shared weights processes grids of any size and shape. Buses are nodes, transmission lines are edges, and the same backbone handles a 500-bus benchmark or a 4,000-bus state-scale topology without per-grid retraining.<\/p>\n\n\n\n<p><strong>Feasibility Aware.<\/strong> Infeasibility is a first-class output, not a discarded label. GridSFM classifies every scenario as feasible or infeasible with a confidence score \u2014 useful for contingency screening, security assessment, and market-clearing validation. On the held-out test set, the classifier reaches <strong>95.3% balanced accuracy<\/strong> (F1 = 0.945 on the feasible class).<\/p>\n\n\n\n<p><strong>Physics Grounded.<\/strong> Branch flows are not predicted directly; they&#8217;re derived analytically from predicted bus voltages and angles via the standard \u03c0-equivalent branch equations. Physics penalties (power balance, thermal, voltage) regularize training so outputs land on the AC-OPF manifold.<\/p>\n\n\n\n<p><strong>Data Efficient.<\/strong> Self-supervised physics constraints supplement supervised solver labels, reducing the per-topology label budget. On a brand-new grid, <strong>as few as ~10 fine-tuning scenarios<\/strong> already produce reasonable cost and dispatch estimates, and <strong>~1,000 scenarios<\/strong> recover full in-sample performance.<\/p>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\" style=\"flex-basis:33.33%\">\n<h2 class=\"wp-block-heading\" id=\"gridsfm-predicts-ac-opf-solutions-in-milliseconds-bus-voltages-generator-dispatch-branch-power-flows-and-a-feasibility-classification-without-running-a-solver\">&#8220;GridSFM predicts AC-OPF solutions in milliseconds: bus voltages, generator dispatch, branch power flows, and a feasibility classification without running a solver.&#8221;<\/h2>\n<\/div>\n<\/div>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"150\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-1024x150.png\" alt=\"GridFM is built around four core tenets: topology-agnostic, feasibility-aware, physics-grounded, and data-efficient.\" class=\"wp-image-1172114\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-1024x150.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-300x44.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-768x112.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-1536x225.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-2048x299.png 2048w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-11-240x35.png 240w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"what-the-model-predicts\">What the model predicts<\/h3>\n\n\n\n<p>Given a grid topology, physical and operating constraints, generation characteristics, and a loading scenario, GridSFM produces a complete operating point \u2014 bus voltage magnitudes V and angles \u03b8, generator active and reactive dispatch (Pg, Qg), branch active and reactive flows (Pij, Qij) \u2014 plus a feasibility verdict with a continuous margin.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"headline-results-gridsfm-open-54-grid-test-corpus\">Headline results (GridSFM-Open, 54-grid test corpus)<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Metric<\/th><th>Value<\/th><\/tr><\/thead><tbody><tr><td>Cost MAPE<\/td><td><strong>3.35%<\/strong> (median 2.85%; 51\/54 grids below 5%)<\/td><\/tr><tr><td>Voltage magnitude MAE<\/td><td>0.0080 p.u.<\/td><\/tr><tr><td>Voltage angle MAE<\/td><td>2.14\u00b0<\/td><\/tr><tr><td>Generator active power MAE<\/td><td>0.092 p.u.<\/td><\/tr><tr><td>Feasibility classifier balanced accuracy<\/td><td><strong>95.3%<\/strong><\/td><\/tr><tr><td>AC-OPF warm-start speedup over cold start<\/td><td><strong>1.66\u00d7 geometric mean<\/strong> (wins on 41\/54 grids)<\/td><\/tr><tr><td>Warm-start speedup vs. DC-OPF baseline<\/td><td>1.59\u00d7 faster than DC warm-start alone<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>Used as a warm-start seed for the PowerModels.jl AC-OPF solver, GridSFM cuts solve time by 1.66\u00d7 on average and captures ~61% of the theoretical headroom between a cold solve and the optimal-point ceiling. On the largest cases this reaches <strong>~4\u00d7 speedup<\/strong> (case1951_rte, case2868_rte) and <strong>6\u20137\u00d7 on a few<\/strong> (Texas2k summer peak, case2742_goc).<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"553\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-12-1024x553.png\" alt=\"Per-grid AC-OPF speedup distribution (log-x axis): GridSFM warm-start vs. DC warm-start vs. ground-truth ceiling, with KDE plots and per-grid dots\" class=\"wp-image-1172117\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-12-1024x553.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-12-300x162.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-12-768x414.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-12-1536x829.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-12-2048x1105.png 2048w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-12-240x130.png 240w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"out-of-distribution-generalization-and-fine-tuning\">Out-of-distribution generalization and fine-tuning<\/h3>\n\n\n\n<p>On a grid 1.4\u00d7 larger than anything seen in training (case6470_rte), zero-shot cost MAPE rises to ~14% \u2014 the model has learned generalizable angle and cost structure, but voltage magnitude and the feasibility classifier need calibration to the new grid. <strong>Fine-tuning on just 1,000 scenarios from the new grid restores accuracy<\/strong> (cost MAPE drops to 1.12%, feasibility F1 recovers to 0.99), and a held-out N-1 contingency split tracks the intact topology closely \u2014 fine-tuning on the base topology transfers cleanly to contingency variants.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"580\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-13-1024x580.png\" alt=\"Fine-tuning loss curves showing rapid adaptation to a previously unseen 6,470-bus grid using only 1,000 training scenarios.\" class=\"wp-image-1172118\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-13-1024x580.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-13-300x170.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-13-768x435.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-13-1536x870.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-13-2048x1160.png 2048w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-13-240x136.png 240w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">Train\/val loss curves over 10 epochs of fine-tuning on case6470_rte (1,000 train graphs)<\/figcaption><\/figure>\n\n\n\n\n\n<h2 class=\"wp-block-heading\" id=\"architecture\">Architecture<\/h2>\n\n\n\n<div class=\"wp-block-columns are-vertically-aligned-top is-layout-flex wp-container-core-columns-is-layout-9d6595d7 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-vertically-aligned-top is-layout-flow wp-block-column-is-layout-flow\" style=\"flex-basis:66.66%\">\n<p>GridSFM is a block-structured discrete neural operator that processes power grids as heterogeneous graphs. Following discrete exterior calculus (DEC) principles, bus quantities (voltage magnitude, angle, nodal injections) are treated as discrete 0-forms on graph vertices, while branch flows are 1-forms on oriented edges. The bus\u2013branch incidence acts as the discrete exterior derivative coupling the two; Kirchhoff&#8217;s current law appears naturally as its codifferential. This formulation gives the architecture a coordinate-free, topology-agnostic view of power flow that transfers across grids of different size and connectivity.<\/p>\n\n\n\n<p>A type-aware projection embeds heterogeneous node and edge features into a shared latent space, augmented with a topology-conditioned learned positional encoding. The latent representation is refined by a stack of N blocks, each applying three sub-operations in sequence with residual skips:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>A <strong>per-type global mixer<\/strong> \u2014 every node attends to every other node of its type.<\/li>\n\n\n\n<li>A <strong>topology-aware mixer<\/strong> along the grid&#8217;s edges \u2014 where node types interact and the grid&#8217;s topology enters.<\/li>\n\n\n\n<li>A <strong>per-type MLP<\/strong>.<\/li>\n<\/ol>\n\n\n\n<p>Five prediction heads output bus voltage magnitude V, angle \u03b8, generator active dispatch Pg, generator reactive dispatch Qg, and a feasibility verdict with a continuous margin. Branch flows (Pij, Qij) are computed analytically from the predicted bus state via the \u03c0-equivalent branch model with off-nominal tap ratio, following the PowerModels.jl polar-AC formulation. This keeps the model on the AC manifold by construction rather than as a soft constraint.<\/p>\n\n\n\n<h3 class=\"wp-block-heading h4\" id=\"training-data\">Training Data<\/h3>\n\n\n\n<p>GridSFM is trained on ~200 base transmission topologies drawn from three open sources:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>PGLib-OPF<\/strong> \u2014 the IEEE PES benchmark library covering networks from hundreds to tens of thousands of buses.<\/li>\n\n\n\n<li><strong>OPFData<\/strong> \u2014 a large-scale AC-OPF dataset derived from PGLib-OPF.<\/li>\n\n\n\n<li><strong>The <code>msr_*<\/code> corpus<\/strong> \u2014 a 48-state continental-US transmission topology set plus six multi-state regions, built by our open-data pipeline (described below).<\/li>\n<\/ul>\n\n\n\n<p>Each base topology is expanded via a <strong>multi-axis perturbation pipeline<\/strong> that varies, independently and in combination:<\/p>\n\n\n\n<p><strong>Generator merit order<\/strong> \u2014 cost coefficients shuffled on 40% of generators, so dispatch learns from physics and cost structure rather than per-grid orderings.<\/p>\n\n\n\n<p><strong>Load profiles<\/strong> \u2014 0.8\u00d7\u20131.5\u00d7 nominal global scaling with \u00b110% per-load jitter; a separate high-load variant from 1.1\u00d7\u20131.3\u00d7 covers the upper-stress tail.<\/p>\n\n\n\n<p><strong>Generator availability<\/strong> \u2014 30% of scenarios contain outages (70% single, 20% double, 10% triple element), weighted by Pmax.<\/p>\n\n\n\n<p><strong>Line ratings<\/strong> \u2014 20% of scenarios derate 10% of branches to 70%\u201395% of nominal.<\/p>\n\n\n\n<p><strong>Voltage limits<\/strong> \u2014 15% of scenarios tighten Vmin\/Vmax on 10% of buses.<\/p>\n\n\n\n<p>A <strong>synthetic-infeasibility pipeline<\/strong> generates targeted failure modes \u2014 voltage squeeze, thermal bottleneck, angle tightening, DC-thermal congestion \u2014 driving a balanced ~50\/50 feasible\/infeasible training mix. Across all topologies and scenarios, training spans <strong>more than half a million labeled samples<\/strong>.<\/p>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-vertically-aligned-top is-layout-flow wp-block-column-is-layout-flow\" style=\"flex-basis:33.33%\">\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"682\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-14-1024x682.png\" alt=\"GridSFM neural architecture: heterogeneous graph embedding feeds N blocks of attention and topological diffusion, with five prediction heads\" class=\"wp-image-1172122\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-14-1024x682.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-14-300x200.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-14-768x511.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-14-1536x1022.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-14-2048x1363.png 2048w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-14-240x160.png 240w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">GridSFM architecture diagram \u2014 heterogeneous-graph embedding, N transformer blocks with topological diffusion and skips, five prediction heads (V, \u03b8, Pg, Qg, feasibility)<\/figcaption><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"on-a-brand-new-grid-as-few-as-10-fine-tuning-scenarios-already-produce-reasonable-cost-and-dispatch-estimates\">&#8220;On a brand-new grid, as few as ~10 fine-tuning scenarios already produce reasonable cost and dispatch estimates.&#8221;<\/h2>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n<\/div>\n<\/div>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"a-pipeline-for-building-realistic-open-data-grid-models\">A pipeline for building realistic open-data grid models<\/h3>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-9d6595d7 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\" style=\"flex-basis:66.66%\">\n<p>Companion to GridSFM, we built a five-stage pipeline that constructs complete, OPF-solvable transmission models entirely from public data:<br><\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Extract<\/strong> power infrastructure from OpenStreetMap via a local Overpass API instance.<\/li>\n\n\n\n<li><strong>Reconstruct<\/strong> bus-branch topology \u2014 voltage inference (neighbor consensus), line merging, and transformer detection.<\/li>\n\n\n\n<li><strong>Estimate<\/strong> electrical parameters using voltage-class lookup tables calibrated with US Energy Information Administration plant-level data.<\/li>\n\n\n\n<li><strong>Allocate<\/strong> hourly demand from EIA-930 to individual buses, using US Census population as a spatial proxy.<\/li>\n\n\n\n<li><strong>Solve<\/strong> both DC and AC optimal power flow using PowerModels.jl with a progressive-relaxation strategy that automatically loosens constraints on imprecise models.<\/li>\n<\/ol>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\" style=\"flex-basis:33.33%\">\n<figure class=\"wp-block-image aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"668\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-15-1024x668.png\" alt=\"Reconstructed Virginia transmission grid: 661 buses, 744 lines, 519 transformers, and 65 generators built entirely from open data sources.\" class=\"wp-image-1172124\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-15-1024x668.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-15-300x196.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-15-768x501.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-15-1536x1002.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-15-240x157.png 240w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-15.png 1864w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">Final bus-branch model for Virginia \u2014 661 buses colored by voltage class, 744 AC lines, 519 transformers, 65 generators sized by capacity<\/figcaption><\/figure>\n<\/div>\n<\/div>\n\n\n\n<p>The pipeline was validated on all 48 contiguous US states and six multi-state regions, including the full Western (5,076 buses) and Eastern (21,697 buses) Interconnections. Of the 48 single-state models:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>42 of 48 (88%)<\/strong> converge at the strictest relaxation level for AC-OPF at peak hour.<\/li>\n\n\n\n<li><strong>44 of 48 (92%)<\/strong> converge at off-peak.<\/li>\n\n\n\n<li><strong>Median dispatch cost: $22\/MWh<\/strong> (range $2.6\/MWh for hydro-dominated Vermont to $104.1\/MWh for import-dependent Rhode Island).<\/li>\n\n\n\n<li><strong>Median system loss: 1.0%<\/strong> (range 0.2%\u20137.1%) \u2014 physically plausible against real US transmission losses of 2\u20133%.<\/li>\n\n\n\n<li><strong>AC\u2013DC cost premium: median 1.8%<\/strong> (0.0\u201313.8%) \u2014 consistent with literature values.<\/li>\n<\/ul>\n\n\n\n<p>All 54 models \u2014 48 single-state and 6 multi-state \u2014 are publicly released at <a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/github.com\/microsoft\/GridSFM\">github.com\/microsoft\/GridSFM<span class=\"sr-only\"> (opens in new tab)<\/span><\/a>.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"comparison-to-prior-approaches\">Comparison to prior approaches<\/h3>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-9d6595d7 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\" style=\"flex-basis:60%\">\n<p>GridSFM is, to our knowledge, the first openly released learned AC-OPF surrogate trained jointly across ~200 base topologies. Most published learned AC-OPF surrogates are <strong>per-grid specialists<\/strong> \u2014 trained and evaluated on a single fixed topology. The closest comparable architecture, gridfm-graphkit (Linux Foundation Energy \/ IBM Research), is in the same accuracy class on per-grid metrics; GridSFM&#8217;s distinguishing properties are (a) a single backbone shared across grids, (b) feasibility as a first-class output, and (c) data efficiency: as few as ~10 fine-tune scenarios yield a useful model on a new grid.<\/p>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\" style=\"flex-basis:40%\">\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"962\" height=\"1024\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-16-962x1024.png\" alt=\"Validated 48-state model results: dispatch costs and fuel mix across the contiguous United States, built from publicly available data only.\" class=\"wp-image-1172125\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-16-962x1024.png 962w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-16-282x300.png 282w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-16-768x817.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-16-1443x1536.png 1443w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-16-169x180.png 169w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/01\/image-16.png 1728w\" sizes=\"auto, (max-width: 962px) 100vw, 962px\" \/><figcaption class=\"wp-element-caption\">States ranked by DC-OPF dispatch cost with installed-capacity fuel mix alongside<\/figcaption><\/figure>\n<\/div>\n<\/div>\n\n\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Small foundation models for the electric grid GridFM is a Microsoft Research initiative to build small foundation models for the electric power grid, applying the same recipe that drives modern language and weather models to the physics of AC power flow. The first release in the family, GridSFM, predicts complete AC Optimal Power Flow (AC-OPF) [&hellip;]<\/p>\n","protected":false},"featured_media":1158679,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","footnotes":""},"research-area":[13556,13548],"msr-locale":[268875],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-1158675","msr-project","type-msr-project","status-publish","has-post-thumbnail","hentry","msr-research-area-artificial-intelligence","msr-research-area-economics","msr-locale-en_us","msr-archive-status-active"],"msr_project_start":"2025-08-01","related-publications":[1170952,1171469],"related-downloads":[],"related-videos":[],"related-groups":[],"related-events":[],"related-opportunities":[],"related-posts":[1170857,1171470],"related-articles":[],"tab-content":[],"slides":[],"related-researchers":[{"type":"user_nicename","display_name":"Andrea Britto Mattos Lima","user_id":42393,"people_section":"Related people","alias":"andreabri"},{"type":"user_nicename","display_name":"Spencer Fowers","user_id":33581,"people_section":"Related people","alias":"sfowers"},{"type":"user_nicename","display_name":"Kate Lytvynets","user_id":38073,"people_section":"Related people","alias":"kalytv"},{"type":"user_nicename","display_name":"Thiago Vallin Spina","user_id":42246,"people_section":"Related people","alias":"tvallinspina"},{"type":"user_nicename","display_name":"Weiwei Yang","user_id":40138,"people_section":"Related people","alias":"weiwya"}],"msr_research_lab":[199565,1161007],"msr_impact_theme":[],"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project\/1158675","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-project"}],"version-history":[{"count":38,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project\/1158675\/revisions"}],"predecessor-version":[{"id":1172138,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-project\/1158675\/revisions\/1172138"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1158679"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1158675"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1158675"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1158675"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1158675"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=1158675"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}