Physics-Aware Transformer-Based Beamforming Optimization for Near-Field E-MIMO ISAC Systems

arXiv:2611.16305 · eess.SP · Submitted 2026-11-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Physics-Aware Transformer-Based Beamforming Optimization for Near-Field E-MIMO ISAC Systems".

Dev: Here's the gist of the paper. This paper proposes a physics-aware Transformer-based beamforming optimization (PATBO) framework for near-field extreme massive multiple-input multiple-output integrated sensing and communication (E-MIMO-ISAC) systems, which is important because it combines structured learning with selective physics-based optimization for low latency and scalable nearfield E-MIMO ISAC beamforming Framework Overview The proposed PATBO framework separates the beamforming problem into two components: physically constructed beam directions and learned stream-power allocation.

Rosa: Rather than directly predicting the complete complex-valued beamforming matrix, the Transformer estimates only the normalized powers assigned to the K communication and L sensing streams.

Dev: For each scenario, PATBO performs four main operations: geometry-aware token construction and Transformer-based power prediction, physical communication and sensing beam-direction generation, beamformer reconstruction and physical feasibility verification, and adaptive physics-based repair and optimizer fallback when required.

Rosa: How the training signal is built The geometry-aware tokens are constructed according to (23)–(25), with feature standardization based exclusively on the training set.

Paper summary: Rosa: This paper, "Physics-Aware Transformer-Based Beamforming Optimization for Near-Field E-MIMO ISAC Systems," introduces the PATBO framework for near-field extreme massive multiple input multiple output integrated sensing and communication systems. The main idea is that they separate the beamforming problem into two parts: physically constructed beam directions and learned stream power allocation.

Dev: What's the core claim of this work, Rosa? What is the thesis here about combining those structured learning aspects with selective physics-based optimization?

Rosa: They propose that instead of predicting the whole complex-valued beamforming matrix directly, which can be huge, the Transformer estimates only the normalized powers assigned to each communication and sensing stream. They build in a way to keep track of which user or target gets which power level during training.

Taro: So they’re using a token-preserving Transformer that is trained with CRB-sensitivity awareness. What does that mean for the complexity of designing these beamformers?

Dev: It means they preserve the link between individual users and targets and their allocated powers, while emphasizing prediction errors having greater impact on near-field localization performance. They are not just predicting a good number; they're predicting a power allocation that respects where you end up in terms of physical constraints.

Rosa: And then after the Transformer makes its guess, there’s a verification step. They check if that predicted beamformer actually respects the physical constraints of the system based on the original operational rules, not just a relaxed constraint for reference generation.

Taro: If it doesn't respect those constraints, what happens? Does it just give up immediately, or is there a way to fix it before moving on?

Dev: They have this hierarchical verify repair fallback strategy. If the prediction is unfeasible, they try to physically constrain and repair the power allocation before resorting to a standard optimizer fallback.

Rosa: The results are pretty strong on feasibility. They say that by using this physics-constrained power repair, they increased joint feasibility from about nineteen point five four percent to eighty seven point seven nine percent on the test set.

Taro: So when you're dealing with these near-field constraints, that jump in feasibility sounds significant. What is the trade-off there?

Dev: They also say that the complete hybrid framework achieves one hundred percent feasibility on the retained test set, and they only needed optimizer fallback for about twelve point two one percent of the samples.

Rosa: And for inference speed, this direct PATBO method is very fast. It takes only one point eleven milliseconds per sample.

Taro: How is that compared to what you're used to? What was the benchmark you were comparing it against?

Dev: They compare it to sequential quadratic programming methods, both cold start and warm start, and PATBO inference takes about eight hundred fifty eight times less time for cold start and seven hundred seventy nine times less time for warm start.

Rosa: It really shows how much faster this can be in real-time applications. They also show that the latency stays around one point one milliseconds even when you scale up to an array size of four hundred to one thousand antennas.

Taro: So it seems like the architecture scales reasonably well for larger systems, which is important when you're dealing with massive antenna arrays and real-world deployment.

Dev: But they do mention a limitation, and this is important for us in the controls side. They point out that while the training works well, numerical optimization stages can still exhibit increasing runtime as the array size gets bigger.

Rosa: That makes sense because those optimization steps are still happening after the initial prediction, even if it's faster than before. So we have a fast prediction but a slower refinement stage depending on how big our system is.

Taro: Does this mean if we go to really massive arrays, say four thousand antennas, the runtime will balloon?

Dev: Yes, they state that the numerical optimization stages exhibit increasing runtime with array size. So for extreme scaling, we still have a latency concern in those later refinement steps.

Rosa: Still, overall their point is that this token-preserving architecture and its CRB-sensitivity training allow them to predict the power allocation in a low-dimensional way while still accounting for the constraints that matter most for localization accuracy.

Taro: So, it's about making the learning process smart enough to care about where we are in terms of physical constraints rather than just getting a generally accurate number.

Dev: That’s exactly what they argue. They emphasize that predicting accurately isn't the same as satisfying those coupled nonconvex constraints, especially when it comes to near-field localization where the Cramer Rao bound depends on things like geometry and distance-dependent steering vector derivatives.

Rosa: So for someone driving or walking, what does this actually mean? Does it change how you think about these massive integrated sensing and communication systems?

Taro: It means we can use AI to do the heavy lifting of power allocation in near-field scenarios much faster than traditional methods, provided we have that verification and repair mechanism in place to handle the physical reality of the system.

Dev: For us on the controls side, it suggests a path toward lower latency operations that are still grounded in physics, even when using deep learning models for the initial setup.

Conclusion: Rosa: So we’re wrapping up this look at the paper "Physics-Aware Transformer-Based Beamforming Optimization for Near-Field E-MIMO ISAC Systems". It’s about how they used a Transformer to predict power in these super complex near-field communication and sensing systems.

Dev: Exactly. They showed that by using that physics-constrained repair instead of just throwing away the prediction when it fails the physical check, you boost your success rate up significantly, from maybe twenty percent up to eighty seven point seven nine percent feasibility on their test set.

Rosa: So the big picture here is that we can use these deep learning models for the initial, fast power allocation guess, and then have physics kick in to make sure it’s actually doable before we commit resources to a full optimization run.

Mohsin Ali, Abdulahi Abiodun Badrudeen, Azharul Islam, Sunwoo Kim

Hanyang University

eess.SP

Submitted: 2026-11-12

Updated: 2026-11-12

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 87/100

The gist: This paper proposes a physics-aware Transformer-based beamforming optimization (PATBO) framework for near-field extreme massive multiple-input multiple-output integrated sensing and communication…

Key concepts

PATBO Framework
This is the proposed method that splits the beamforming problem into two parts: physically constructed beam directions and learned stream-power allocation. Instead of calculating all complex coefficients, the Transformer estimates only the normalized powers for communication and sensing streams.
Geometry-Aware Tokens
These are specific data representations created for each communication user and sensing target. They encode geometric information like coordinates, angles, and reference parameters into a structured format that the Transformer can process to understand spatial relationships.
CRB-Sensitivity-Aware Training
This is a training technique where the neural network is trained to focus its prediction errors on stream powers that most significantly affect the localization performance (CRB). This ensures the learned power allocation is optimized for accurate sensing results.

Terminology

Summary

This paper proposes a physics-aware Transformer-based beamforming optimization (PATBO) framework for near-field extreme massive multiple-input multiple-output integrated sensing and communication (E-MIMO-ISAC) systems, which is important because it combines structured learning with selective physics-based optimization for low latency and scalable nearfield E-MIMO ISAC beamforming

Framework Overview

The proposed PATBO framework separates the beamforming problem into two components: physically constructed beam directions and learned stream-power allocation. Rather than directly predicting the complete complex-valued beamforming matrix, the Transformer estimates only the normalized powers assigned to the K communication and L sensing streams. p. 3

Where the paper says it

Overview The proposed PATBO framework separates the beamforming problem into two components: physically constructed beam directions and learned stream-power allocation. Rather than directly predicting the complete complex-valued beamforming matrix, the Transformer estimates only the normalized powers assigned to the K communication and L sensing streams.

For each scenario, PATBO performs four main operations: geometry-aware token construction and Transformer-based power prediction, physical communication and sensing beam-direction generation, beamformer reconstruction and physical feasibility verification, and adaptive physics-based repair and optimizer fallback when required. p. 3
Where the paper says it

Rather than directly predicting the complete complex-valued beamforming matrix, the Transformer estimates only the normalized powers assigned to the K communication and L sensing streams. For each scenario, PATBO performs four main operations: 1) geometry-aware token construction and Transformerbased power prediction; 2) physical communication and sensing beam-direction generation; 3) beamformer reconstruction and physical feasibility verification; and 4) adaptive physics-based repair and optimizer fallback when required.

How the training signal is built

The geometry-aware tokens are constructed according to (23)–(25), with feature standardization based exclusively on the training set. Each communication user and sensing target is represented by an individual Transformer token, such as zc k = [xk, yk, zk, rk, θk, ϕk, 0, 0, 0, 0]T. p. 4

Where the paper says it

Geometry-Aware Token Representation Each communication user and sensing target is represented by an individual Transformer token. The communication-user token is defined as z c k = [xk, yk, zk, rk, θk, ϕk, 0, 0, 0, 0]T , (23) where (xk, yk, zk) denotes the Cartesian position, rk denotes the distance from the BS reference point, and θk and ϕk denote the corresponding angular coordinates.

The latent sequence is processed by two Transformer encoder stages each employing four attention heads and a feedforward dimension of 128. p. 4
Where the paper says it

, z (0) K+L i ∈ R dmodel×(K+L) (27) denote the embedded token sequence. The latent sequence is processed by two Transformer encoder stages, each employing four attention heads and a feedforward dimension of 128.

Through self-attention, each token incorporates information from all communication users and sensing targets. p. 4
Where the paper says it

, z (T) K+L i . (28) Through self-attention, each token incorporates information from all communication users and sensing targets.

The resulting K + L entity-specific logits are jointly normalized through a softmax operation to yield the predicted normalized stream-power vector ρb. p. 5
Where the paper says it

The resulting K + L entity-specific logits are jointly normalized through a softmax operation: ρˆq = exp(ℓq) K X +L j=1 exp(ℓj ) , q = 1, . (30) Therefore, the predicted normalized stream-power vector is ρb = [ˆρ1, .

Where the gains come from

The main contributions of this work are summarized as follows:

  1. We develop a structured near-field E-MIMO-ISAC beamforming framework that combines RZF communication directions and spherical-wave near-field sensing directions with learned stream-power allocation. p. 1

    Where the paper says it

    After inference, physical feasibility is explicitly verified; infeasible predictions are passed to a low-dimensional power-repair stage, and optimizer fallback is invoked when repair is unsuccessful. The main contributions of this work are summarized as follows: • We develop a structured near-field E-MIMO-ISAC beamforming framework that combines RZF communication directions and spherical-wave near-field sensing directions with learned stream-power allocation.

    For M = 1024, K = 4, and L = 2, this reduces the learned output representation from 12,288 real-valued beamforming coefficients to only six stream-power variables. p. 1
    Where the paper says it

    The main contributions of this work are summarized as follows: • We develop a structured near-field E-MIMO-ISAC beamforming framework that combines RZF communication directions and spherical-wave near-field sensing directions with learned stream-power allocation. For M = 1024, K = 4, and L = 2, this reduces the learned output representation from 12,288 real-valued beamforming coefficients to only six stream-power variables.

  2. We propose a token-preserving Transformer with CRB-sensitivity-aware training that preserves the association between individual users/targets and their allocated powers while emphasizing prediction errors having greater impact on near-field localization performance. p. 1

    Where the paper says it

    • We propose a token-preserving Transformer with CRBsensitivity-aware training. The architecture preserves the association between individual users/targets and their allocated powers, while emphasizing prediction errors having greater impact on near-field localization performance.

  3. We develop a hierarchical verify–repair–fallback strategy for physical constraint satisfaction, where physics-constrained power repair increases joint feasibility from 19. p. 1

    Where the paper says it

    • We develop a hierarchical verify–repair–fallback strategy for physical constraint satisfaction. Physics-constrained power repair increases joint feasibility from 19.

    54% to 87.79%, while the complete hybrid framework achieves 100% feasibility on the retained test set with optimizer fallback required for only 12. p. 1
    Where the paper says it

    Physics-constrained power repair increases joint feasibility from 19. 79%, while the complete hybrid framework achieves 100% feasibility on the retained test set with optimizer fallback required for only 12.

    21% of the samples. p. 1
    Where the paper says it

    79%, while the complete hybrid framework achieves 100% feasibility on the retained test set with optimizer fallback required for only 12. 21% of the samples.

Constraint Handling and Verification

The final PATBO formulation deliberately employs only the CRB-sensitivity weighting in the neural loss, assigning larger penalties to prediction errors associated with streams having stronger local impact on the aggregate localization CRB. After reconstructing the Transformer-predicted beamformer Wc according to (38), its feasibility is evaluated using the original operational constraints rather than the tightened CRB threshold used for reference generation. p. 6

Where the paper says it

The complete CRB-sensitivity estimation and PATBO training procedure is summarized in Algorithm 2. Physical Feasibility Verification After reconstructing the Transformer-predicted beamformer Wc according to (38), its feasibility is evaluated using the original operational constraints rather than the tightened CRB threshold used for reference generation.

If F(Wc) = 1, the Transformer output is directly accepted; otherwise, physics-constrained repair is invoked. p. 6
Where the paper says it

If F(Wc) = 1, the Transformer output is directly accepted; otherwise, physicsconstrained repair is invoked. Physics-Constrained Power Repair If the Transformer prediction violates one or more original constraints, PATBO invokes a low-dimensional physicsconstrained power-repair stage.

Performance and Scalability

The complete hybrid PATBO framework achieves an average sum rate of 17.9004 bit/s/Hz, corresponding to 99. p. 9

Where the paper says it

The final hybrid beamformer achieves an average sum rate of 17. 9004 bit/s/Hz, corresponding to 99.

89% reference-rate retention. p. 0
Where the paper says it

Direct PATBO achieves 100. 98% reference-rate retention and 19.

Direct PATBO inference requires only 1. p. 0
Where the paper says it

21% of the scenarios. Direct PATBO inference requires 1.

p. 0
Where the paper says it

21% of the scenarios. Direct PATBO inference requires 1.

11 ms per sample, corresponding to approximately 858× and 779× lower runtime than the evaluated cold-start and PATBO-warm-start sequential quadratic programming (SQP) implementations, respectively. p. 0
Where the paper says it

Direct PATBO inference requires 1. 11 ms per sample, corresponding to approximately 858× and 779× lower runtime than the evaluated cold-start and PATBO-warm-start sequential quadratic programming (SQP) implementations, respectively.

The scalability results show that PATBO maintains approximately 1. p. 0
Where the paper says it

Direct PATBO inference requires 1. 11 ms per sample, corresponding to approximately 858× and 779× lower runtime than the evaluated cold-start and PATBO-warm-start sequential quadratic programming (SQP) implementations, respectively.

1 ms inference latency across M = 400–1024, while the numerical optimization stages exhibit increasing runtime with array size. p. 7
Where the paper says it

Direct Transformer inference offers the lowest online latency, whereas physics-constrained repair and reference optimization can be invoked when stronger feasibility guarantees are required. An additional advantage of the structured formulation is its favorable scaling with the array size.

p. 7
Where the paper says it

Direct Transformer inference offers the lowest online latency, whereas physics-constrained repair and reference optimization can be invoked when stronger feasibility guarantees are required. An additional advantage of the structured formulation is its favorable scaling with the array size.

Conclusion

The proposed token-preserving architecture and CRB-sensitivity-aware training enable low-dimensional power prediction while accounting for localization-sensitive constraints. Numerical results show that physics-constrained repair increases joint feasibility from 19. p. 1

Where the paper says it

• We develop a hierarchical verify–repair–fallback strategy for physical constraint satisfaction. Physics-constrained power repair increases joint feasibility from 19.

54% for direct PATBO to 87. p. 0
Where the paper says it

For a 32 × 32 E-MIMO array with 1024 antennas, the proposed structured representation reduces the learned output from 12,288 real-valued coefficients of an unstructured complex beamformer to six stream-power variables. Direct PATBO achieves 100.

79%, while the complete verify–repair–fallback framework achieves 99. p. 1
Where the paper says it

Physics-constrained power repair increases joint feasibility from 19. 79%, while the complete hybrid framework achieves 100% feasibility on the retained test set with optimizer fallback required for only 12.

89% reference-rate retention and 100% feasibility on the retained test set. p. 0
Where the paper says it

The complete hybrid framework achieves 99. 89% reference-rate retention and 100% constraint satisfaction on the retained test set, while requiring optimizer fallback for only 12.

Direct PATBO inference requires only 1.11 ms per sample, corresponding to approximately 858× and 779× lower runtime than the evaluated coldstart and PATBO-warm-start sequential quadratic programming (SQP) benchmarks. p. 0
Where the paper says it

Direct PATBO inference requires 1. 11 ms per sample, corresponding to approximately 858× and 779× lower runtime than the evaluated cold-start and PATBO-warm-start sequential quadratic programming (SQP) implementations, respectively.

Moreover, the scalability results show that PATBO maintains approximately 1. p. 0
Where the paper says it

Direct PATBO inference requires 1. 11 ms per sample, corresponding to approximately 858× and 779× lower runtime than the evaluated cold-start and PATBO-warm-start sequential quadratic programming (SQP) implementations, respectively.

1 ms inference latency across M = 400–1024, while the numerical optimization stages exhibit increasing runtime with array size. The results demonstrate the potential of combining structured learning with selective physics-based optimization for low-latency and scalable nearfield E-MIMO-ISAC beamforming. p. 1
Where the paper says it

Such a structured representation reduces the learning dimension while retaining the near-field spatial characteristics through physics-based beam construction. To address these challenges, this paper proposes a physicsaware Transformer-based beamforming optimization (PATBO) framework for near-field E-MIMO-ISAC systems.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Physics-aware Transformer-based beamforming optimization (PATBO) allows AI systems to achieve low-latency inference by predicting only K + L normalized stream powers instead of 2M real-valued coefficients, leading to a 2048x reduction in learned output dimensionality for an M=1024 array.

  2. The token-preserving Transformer architecture maintains the correspondence between individual users/targets and their allocated powers, ensuring that each output power corresponds to a specific physical entity during prediction.

  3. The CRB-sensitivity-aware training objective assigns larger penalties to prediction errors associated with streams having stronger local impact on the aggregate localization CRB by defining a stream-wise prediction error as eq,n = ˆρq,n − ρ ⋆ q,n.

  4. A hierarchical verify–repair–fallback strategy enables the AI system to achieve high joint feasibility by explicitly verifying physical constraints after inference and invoking physics-constrained power repair or optimizer fallback when necessary.

  5. The improved AI system can handle near-field E-MIMO integrated sensing and communication (E-MIMO-ISAC) systems with 100% constraint satisfaction on the retained test set, which is essential for reliable deployment in complex near-field environments.

Abstract

This paper proposes a physics-aware Transformer-based beamforming optimization (PATBO) framework for near-field extreme massive multiple-input multiple-output integrated sensing and communication (E-MIMO-ISAC) systems. PATBO combines regularized zero-forcing (RZF) communication directions and spherical-wave near-field sensing directions with learned stream-power allocation. A token-preserving Transformer maintains the correspondence between individual users/targets and their allocated powers, while a Cramér--Rao bound (CRB)-sensitivity-aware objective emphasizes power-allocation errors with greater localization impact. Physical feasibility is explicitly verified after inference, with infeasible predictions handled through physics-constrained power repair and optimizer fallback. For a 32 times32 E-MIMO array with 1024 antennas, the proposed structured representation reduces the learned output from 12, 288 real-valued coefficients of an unstructured complex beamformer to six stream-power variables. Direct PATBO achieves 100.98% reference-rate retention and 19.54% joint feasibility, which increases to 87.79% after repair. The complete hybrid framework achieves 99.89% reference-rate retention and 100% constraint satisfaction on the retained test set, while requiring optimizer fallback for only 12.21% of the scenarios. Direct PATBO inference requires 1.11 ms per sample, corresponding to approximately 858 times and 779 times lower runtime than the evaluated cold-start and PATBO-warm-start sequential quadratic programming (SQP) implementations, respectively.

Related papers