arxiv: 2605.05952 · v1 · submitted 2026-05-07 · 📡 eess.SY · cs.SY

Recognition: unknown

Foundation Twins: A New Generation of Power Systems Digital Twins using Foundation AI Models

Pedro P. Vergara

Authors on Pith no claims yet

Pith reviewed 2026-05-08 07:03 UTC · model grok-4.3

classification 📡 eess.SY cs.SY

keywords foundation twinspower systemsdigital twinsfoundation modelsreinforcement learningartificial intelligenceenergy systemsmulti-timescale modeling

0 comments

The pith

Foundation Twins combine foundation models with reinforcement learning to create versatile power system digital twins.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper is a position paper that envisions a new type of digital twin for power systems called Foundation Twins. It argues that these twins can be created by merging the broad generalization capabilities of foundation AI models with the decision-making strengths of reinforcement learning. This matters because power systems involve complex phenomena and decisions that occur at different speeds and over different areas, and current digital twins fall short in handling all of them together. If successful, this approach would allow for faster and better informed choices in operating and planning energy networks. The author believes that recent progress in AI makes it possible to build these integrated systems now.

Core claim

The author proposes Foundation Twins as the way to realize practical power systems digital twins. By combining the generalization features of foundation models with the decision-making capabilities of reinforcement learning architectures, these twins can handle the multi-timescale and multi-scope nature of power systems. This would allow modeling and simulation tools to accelerate and improve decision-making across different time scales and geographic scopes.

What carries the argument

Foundation Twins, the proposed integration of foundation models for broad generalization and reinforcement learning architectures for decision-making in multi-timescale power system simulations.

If this is right

Power systems digital twins would support decisions spanning short-term operations to long-term planning.
The twins would address both local and wide geographic scopes in energy systems.
Digital twins would move from conceptual ideas to practical implementation tools.
Decision-making processes in power systems would become more unified across time horizons.
Research would shift toward developing hybrid AI architectures for infrastructure modeling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

One could test this by creating a prototype that adapts a foundation model to power data and trains it with RL for control tasks.
This vision might apply to other domains with multi-scale dynamics, such as transportation or climate systems.
If it works, it could reduce the need for many separate models by having one system handle multiple tasks.
Extensions could include adding physics-based constraints to improve accuracy in power flow calculations.

Load-bearing premise

Recent advances in foundation models and reinforcement learning can be combined in practice to create digital twins that effectively address the multi-timescale, multi-scope nature of power systems.

What would settle it

Building a Foundation Twin prototype and observing whether it can reliably predict or optimize power system behavior over both millisecond-level transients and yearly planning periods would determine if the combined approach succeeds.

Figures

Figures reproduced from arXiv: 2605.05952 by Pedro P. Vergara.

**Figure 1.** Figure 1: Approximated power systems processes and their diffe view at source ↗

**Figure 2.** Figure 2: Foundation Twins architecture composed of several fo view at source ↗

**Figure 3.** Figure 3: Representation of a hierarchical learning problem w view at source ↗

read the original abstract

Power systems are inherently multi-timescale systems, with different physical phenomena and decision-making processes spanning multiple timescales, time horizons, and geographic scopes. I envision power systems digital twins (DTs) as powerful modeling and simulation tools that can accelerate and improve decision-making across different time scales and geographic scopes. However, until now, research has not delivered such a vision, and power systems DTs remain a concept distant from implementation. This is not a regular research paper. This is a position paper that outlines my vision for developing a new generation of power systems DTs that leverage recent advances in artificial intelligence (AI) and machine learning (ML). I call these Foundation Twins. Foundation Twins combines the generalization features of foundation models with the decision-making capabilities of reinforcement learning (RL) architectures to deliver the envisioned power systems DTs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This position paper sketches a vision for 'Foundation Twins' in power systems but stays entirely conceptual with no technical details, experiments, or evidence.

read the letter

The key takeaway is that this is a position paper laying out a vision for power systems digital twins based on foundation AI models and reinforcement learning, without any implemented work or supporting experiments. It does a decent job of describing the core challenge: power systems operate across many timescales and geographic scopes, from fast transients to long-term planning. The author notes that existing digital twins haven't reached the level of integration needed for real decision support, especially with renewables coming in. The main novelty is framing the combination of foundation models' broad generalization with RL's decision-making as 'Foundation Twins.' This builds on existing ideas in AI but applies them specifically to this infrastructure problem. Where it falls short is the complete lack of substance beyond the concept. There are no details on model architectures, how to handle the different timescales in practice, data requirements, or potential pitfalls like computational cost or reliability. The assumption that these AI tools can be combined effectively to solve the multi-timescale issue is stated but not explored or evidenced in any way. Because it's purely forward-looking, the paper doesn't advance the technical state of the art. Readers looking for concrete methods or results won't find them here. This would be useful for a reading group focused on emerging applications of large AI models in engineering domains. It could help generate ideas for actual research projects. I wouldn't cite the paper in my own work since it doesn't contain verifiable contributions. However, it deserves to go through peer review so that experts in power systems and AI can comment on whether the proposed direction is promising or if there are obvious barriers the vision overlooks. A serious editor should send it out rather than desk reject it.

Referee Report

0 major / 2 minor

Summary. The manuscript is a position paper proposing 'Foundation Twins' as a new generation of power systems digital twins. It envisions combining the generalization capabilities of foundation AI models with the decision-making strengths of reinforcement learning architectures to address the multi-timescale, multi-scope, and multi-horizon challenges inherent in power systems modeling, simulation, and control.

Significance. If the proposed integration of foundation models and RL can be realized, the vision could substantially advance power systems digital twins by enabling more adaptive, generalizable tools that span transient dynamics to long-term planning across geographic scales, potentially improving operational decision-making and system resilience.

minor comments (2)

[Abstract] The abstract and description would benefit from one or two concrete examples of multi-timescale phenomena (e.g., electromagnetic transients versus unit commitment) to ground the stated requirements.
A short discussion of related work on existing power-systems digital twins or early applications of foundation models in engineering domains would help situate the novelty of the Foundation Twins proposal.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive review and recommendation to accept the manuscript. We appreciate the recognition that this is a position paper outlining a vision for Foundation Twins that combines foundation models with reinforcement learning to address multi-timescale challenges in power systems.

Circularity Check

0 steps flagged

No circularity: purely conceptual position paper with no derivations or equations

full rationale

The manuscript explicitly identifies itself as a position paper that articulates a high-level vision for 'Foundation Twins' rather than presenting completed research, empirical results, or a concrete architecture with equations. No load-bearing steps exist that reduce by construction to inputs, fitted parameters, or self-citations, as there are no mathematical derivations, predictions, or technical claims to inspect. The central proposal—that foundation models' generalization combined with RL can address multi-timescale power systems DTs—functions as a forward-looking conceptual outline without internal reductions.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that power systems are multi-timescale and that foundation models plus RL can address this without specific mechanisms or evidence provided.

axioms (1)

domain assumption Power systems are inherently multi-timescale systems, with different physical phenomena and decision-making processes spanning multiple timescales, time horizons, and geographic scopes.
Invoked in the abstract as the foundational premise for needing new DTs.

invented entities (1)

Foundation Twins no independent evidence
purpose: New generation of power systems digital twins using foundation AI models combined with RL.
Coined term for the proposed architecture; no independent evidence or implementation details given.

pith-pipeline@v0.9.0 · 5434 in / 1195 out tokens · 51608 ms · 2026-05-08T07:03:55.565072+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 2 canonical work pages · 1 internal anchor

[1]

On future power system digital twins: A vision towards a standard arch itecture,

W. Zomerdijk, P . Palensky, T. AlSkaif, and P . P . V ergara, “On future power system digital twins: A vision towards a standard arch itecture,” Digital Twins and Applications , vol. 1, no. 2, p. 103–117, Dec. 2024

2024
[2]

Milano, Power System Modelling and Scripting

F. Milano, Power System Modelling and Scripting . Springer Berlin Heidelberg, 2010

2010
[3]

Foundation models for the electric power gr id,

H. F. Hamann, B. Gjorgiev, T. Brunschwiler, L. S. Martins , A. Puech, A. V arbella, J. Weiss, J. Bernabe-Moreno, A. B. Mass´ e, S. L. Choi, I. Foster, B.-M. Hodge, R. Jain, K. Kim, V . Mai, F. Mirall` es,M. De Mon- tigny, O. Ramos-Lea˜ nos, H. Suprˆ eme, L. Xie, E.-N. S. Y ouss ef, A. Zinﬂou, A. Belyi, R. J. Bessa, B. P . Bhattarai, J. Schmude, and S. Sob...

2024
[4]

A path towards autonomous machine intelli- gence,

Y . LeCun, “A path towards autonomous machine intelli- gence,” OpenReview, 2022, position paper. [Online]. Available: https://openreview.net/forum?id=BZ5a1r-kVsf

2022
[5]

Comparative assessment of generative models for transformer- and consumer-level load proﬁles generation,

W. Xia, H. Huang, E. M. S. Duque, S. Hou, P . Palensky, and P . P . V ergara, “Comparative assessment of generative models for transformer- and consumer-level load proﬁles generation,” Sustainable Energy, Grids and Networks , vol. 38, p. 101338, Jun. 2024

2024
[6]

Pow erpm: Foundation model for power systems,

S. Tu, Y . Zhang, J. Zhang, Z. Fu, Y . Zhang, and Y . Y ang, “Pow erpm: Foundation model for power systems,” in Advances in Neural Informa- tion Processing Systems , vol. 37. Curran Associates, Inc., 2024, pp. 115 233–115 260

2024
[7]

Energydiff: Universal time-series energy data generation using diffusion models,

N. Lin, P . Palensky, and P . P . V ergara, “Energydiff: Universal time-series energy data generation using diffusion models,” IEEE Transactions on Smart Grid , vol. 16, no. 5, pp. 4252–4265, 2025

2025
[8]

A ﬂow-based model for conditional and probabilistic electricity consumption pr oﬁle generation,

W. Xia, C. Wang, P . Palensky, and P . P . V ergara, “A ﬂow-based model for conditional and probabilistic electricity consumption pr oﬁle generation,” Energy and AI , vol. 21, p. 100586, Sep. 2025

2025
[9]

Smart- MeterFM: Unifying smart meter data generative tasks using ﬂ ow match- ing models,

N. Lin, Y . Wang, J. Heres, P . Palensky, and P . P . V ergara, “ Smart- MeterFM: Unifying smart meter data generative tasks using ﬂ ow match- ing models,” 10.48550/arXiv.2601.21706, 2026

work page doi:10.48550/arxiv.2601.21706 2026
[10]

Physics-informed machine learning,

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P . Perdikari s, S. Wang, and L. Y ang, “Physics-informed machine learning,” Nature Reviews Physics, vol. 3, no. 6, p. 422–440, May 2021

2021
[11]

Physics informe d neural net- works for continuum micromechanics,

A. Henkes, H. Wessels, and R. Mahnken, “Physics informe d neural net- works for continuum micromechanics,” Computer Methods in Applied Mechanics and Engineering , vol. 393, p. 114790, Apr. 2022

2022
[14]

Powerﬂownet: Power ﬂow approximation using message passi ng graph neural networks,

N. Lin, S. Orfanoudakis, N. O. Cardenas, J. S. Giraldo, a nd P . P . V ergara, “Powerﬂownet: Power ﬂow approximation using message passi ng graph neural networks,” Int. J. of Electrical Power and Energy Systems , vol. 160, p. 110112, Sep. 2024

2024
[15]

Deep hi erarchical planning from pixels,

D. Hafner, K.-H. Lee, I. Fischer, and P . Abbeel, “Deep hi erarchical planning from pixels,” in Advances in Neural Information Processing Systems, vol. 35, 2022, pp. 26 091–26 104

2022
[16]

SAVGO: Learning State-Action Value Geometry with Cosine Similarity for Continuous Control

S. Orfanoudakis and P . P . V ergara, “SA VGO: Learning sta te-action value geometry with cosine similarity for continuous contr ol,” 10.48550/ARXIV .2605.00787, 2026

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2026
[17]

Introduction to latent variable energy-based models: a path toward autonomous machine intelligence,

A. Dawid and Y . LeCun, “Introduction to latent variable energy-based models: a path toward autonomous machine intelligence,” Journal of Statistical Mechanics: Theory and Experiment , vol. 2024, no. 10, p. 104011, Oct. 2024

2024