Papers
Machine learning for physics, as it is published.
New work from arXiv on CFD surrogates, neural operators and physics-informed learning, refreshed every hour. Search it, filter by topic, or start with the papers we keep coming back to.
Papers
New work from arXiv on CFD surrogates, neural operators and physics-informed learning, refreshed every hour. Search it, filter by topic, or start with the papers we keep coming back to.
1,058 papers in CFD surrogates · newest first
physics.optics
Coarse-mode aerosols (CMAs), including pollen, spores, and dust, remain difficult to characterize in situ because conventional instruments rarely resolve particle geometry and number concentration together. We present HoloCMA, which combines digital inline holography, numerical reconstruction, and deep learning-assisted analysis for particle-resolved measurement. Its optical system targets particles from approximately 5.0 um to the millimeter scale. HoloCMA supports active sampling at up to 30.0 L/min and converts to open-path operation by removing the sampling module. With an NVIDIA GeForce RTX 5070 Laptop GPU and 8 GB VRAM, HoloCMA reports equivalent circular diameter (ECD) and number concentration in real time without sustained queue buildup up to 11.4 particles/cm3. Additional computing could extend real-time analysis to the workflow limit of 100.0 particles/cm3. We evaluated HoloCMA using polystyrene latex spheres, sodium chloride and ammonium sulfate crystals, and oleic acid droplets with nominal diameters of 4.9-11.8 um, with an Aerodynamic Particle Sizer (APS) for comparison. Across 12 comparisons, absolute differences between HoloCMA and corrected APS geometric-mean diameters ranged from less than 0.1 to 1.0 um, with a mean absolute difference of 0.5 um and a mean absolute relative difference of 8.4%. HoloCMA measures geometric size directly from reconstructed contours, whereas converting APS aerodynamic diameter requires material-specific assumptions. The APS-to-HoloCMA concentration ratio generally decreased with particle size, consistent with size-dependent transport and counting losses in the APS, although the cause could not be determined conclusively. HoloCMA thus enables continuous CMA measurements that combine geometric size, number concentration, and retained particle images, supporting classification and long-term atmospheric, environmental, and indoor-air monitoring.
stat.ML
We address the simultaneous prediction of multiple high-dimensional physical fields governed by linear equality constraints, a setting that arises in many real-world applications in physics machine learning. Gaussian process (GP) regression is a widely used surrogate modeling approach due to its effectiveness in small-sample regimes and its ability to provide uncertainty quantification. However, applying GP models in this setting raises two major challenges: the high dimensionality of the discretized output fields and the enforcement of the physical constraint in predictions. For the latter, a common strategy consists in deducing one output from the others via the constraint relation. Through a benchmark, we show that this deductive approach is sensitive to the arbitrary choice of which output to deduce, affecting both predictive accuracy and uncertainty quantification. Consequently, there is a need for an approach that treats all fields symmetrically while strictly respecting the underlying physics. Motivated by these limitations, we propose a robust framework for jointly modeling constrained multi-field data. Our approach first leverages a specific PCA procedure for multi-field data, coined row-wise PCA, which has the interesting property of preserving the constraint in the latent space. Since standard PCA strategies for multi-field data do not preserve such constraints, we investigate theoretically the optimality of the row-wise choice. In a second step, we consider a linearly-constrained multi-output GP approach based on a specific kernel parametrization which is trained on the latent space of row-wise PCA. The proposed framework is validated on a population dynamics problem and on an industrial CFD application, which involves the prediction of Reynolds stress tensor components under the incompressibility constraint.
physics.ed-ph
FlowMLLab is an open-source framework that connects validated fluid-mechanics solvers to scientific machine-learning experiments under a reproducible contract. Its computational fluid dynamics (CFD) pathway generates cavity data, whereas its direct simulation Monte Carlo (DSMC) pathway supplies rarefied-flow evidence with sampling diagnostics. Case-wise partitions prevent leakage, and matched non-neural baselines determine whether learning offers a measurable advantage. Examples include a coordinate network and a proper orthogonal decomposition (POD) deep operator network (DeepONet), together with learned kinetic closures. Python components regenerate every figure, while sixteen tutorials use the same interfaces as the command-line tools. By retaining numerical provenance, blind-case protocols and physical validation targets, FlowMLLab emphasizes traceable evidence rather than black-box curve fitting.
cs.AI
Neural operators have emerged as efficient surrogates for time-dependent physical systems governed by partial differential equations (PDEs), but their future-state predictions are often conditioned only on observed states and static problem descriptors. For control or optimization, however, body motion, inflow, or forcing are prescribed for the query without being determined solely by the observed state. We introduce the Spatiotemporal Conditional Operator (STCO) for prescribed-condition operator learning (PCOL), a common interface that supplies prescribed target-time condition fields to heterogeneous backbone architectures while retaining their architecture-specific core computation and context pathways. Its condition interface combines Flow-Aware Graph Leaf (FAGL) with Dual-Site Feature-wise Linear Modulation (DSFiLM). Non-learned FAGL uses vorticity from the final observed frame to construct a fixed-cardinality adaptive partition, then co-locates the observed history and target-time condition fields at its regional coordinates. DSFiLM injects separate motion, inflow, and force routes before and after operator computation through current-feature-driven slot- and channel-wise gates. We evaluate twelve matched backbone architectures with different existing physical and temporal inputs. The immersed-boundary computational fluid dynamics (CFD) benchmark spans prescribed motion, inflow disturbances, body-force actuation, and morphology. Across twelve matched backbones, three regimes, and two lead ranges, STCO yields mean paired reductions of 31.1% in relative-L2 field error and 24.7% in normalized pressure-derived load error. It also lowers longer-lead field error for 11 backbones, while interventions on individual condition groups produce measurable prediction changes for every group evaluated.
cs.LG
Physics-informed neural networks (PINNs) embed governing partial differential equations directly into the training loss, offering a promising alternative to costly CFD solvers for unsteady flows. Yet the growing list of techniques proposed to improve PINN training is typically validated one at a time, leaving open whether these techniques actually compose. We study this question in depth on the DFG/Schafer-Turek unsteady cylinder wake benchmark. In isolation, nearly every technique performs no better than an untreated baseline. However, combining periodic (SIREN) activations with causal weighting unlocks a previously inaccessible regime, reconstructing velocity and pressure fields to within 4.1% average relative L2 error against an OpenFOAM reference solution. Adding further techniques instead causes catastrophic performance degradation, demonstrating that individually effective PINN interventions can interact nonlinearly and that more elaborate training recipes are not necessarily better.
physics.flu-dyn
Granular flows are ubiquitous in natural and industrial systems, yet their complex dynamics remain difficult to characterize. For inverse problems involving unknown inlet, outlet, and wall boundary conditions, where CFD simulations are challenging, reconstructing complete flow fields from sparse observations constitutes a challenging inverse problem. In this study, a physics-informed neural network framework driven by both physical mechanisms and measurement data is developed to reconstruct the steady-state full-field distribution of granular flows in a pipe. The proposed approach integrates sparse measurement data with governing equations and constitutive relations and is trained using high-fidelity datasets generated by CFD solutions of a continuum model. The framework incorporates a dimensionless loss formulation, physics-informed initialization, dynamic global weighting, and a locally weighted granular temperature data-loss strategy. These treatments enable accurate reconstruction of the complete flow-field evolution. This work establishes a robust methodological framework for flow-field reconstruction in complex granular flow systems.
physics.flu-dyn
This paper investigates the feasibility of using residual learning to improve unsteady aerodynamic load prediction for aeroelastic applications. The machine learning technique selected for the study is the long short-term memory (LSTM) neural network, which is used for its suitability for sequential data with aerodynamic memory effects. The approach is investigated for the NLR 7301 airfoil benchmark using high-fidelity CFD lift data for prescribed pitch and plunge motions in the transonic flow regime in the presence of shock motion. An analytical unsteady aerodynamic model based on the Wagner function is used as a physics-based baseline, and the neural network is trained to learn the difference between the CFD lift coefficient and the Wagner prediction. The residual model is compared with a direct neural-network model trained to predict the CFD lift coefficient. The comparison includes feature and normalization studies, external benchmark cases, and leave-one-out and leave-family-out generalization tests across a range of sinusoidal and non-sinusoidal motions. The residual model performs best when its inputs align with the Wagner formulation variables, generally giving lower error and more consistent performance across training runs, though the direct model remains more accurate for some high-frequency cases. The residual model also generalizes better in the leave-one-out and leave-family-out tests, with a smaller increase in error than the direct model when entire motion families are withheld from training. Overall, the results indicate that residual learning shows promise as a modular approach for augmenting classical low-order aerodynamic theories, especially when the physics baseline removes a structured part of the aerodynamic response and leaves a lower-variance correction for the neural network to learn.
physics.flu-dyn
Spray atomisation rapidly creates large liquid-gas interfacial areas and is central to many industrial processes. However, predicting spray behaviour and surface area remains difficult: experiments cannot access all spray regions, while CFD becomes prohibitively expensive as finer structures develop. Data driven surrogates can learn interface evolution, enabling rapid design space exploration, operating condition ranking, and ultimately spray control. We investigate how state representation, neural architecture, and physics-informed regularisation affect long horizon autoregressive forecasting of spray interfaces, particularly conservation. Our principal model is a boundary-conditioned Fourier Neural Operator (FNO) that predicts the evolution of the signed distance function (SDF) from the liquid-gas interface. It is trained on 2D sharp interface Volume-of-Fluid CFD simulations spanning several atomisation regimes. The SDF-FNO retains interface fidelity better than an FNO trained directly on volume fraction, but is outperformed by a U-Net. Objective function ablation shows that a liquid inventory penalty improves conservation at a modest cost to local interface accuracy. We also introduce a physics-informed extension combining an open-domain target-increment liquid balance penalty, a narrowband Eikonal regulariser that preserves signed distance geometry, and a phase-boundedness penalty. Although this model trains stably, it leaves forecast error, interface overlap, and inventory behaviour essentially unchanged relative to the data driven baseline. Finally, we demonstrate the surrogate by ranking injection conditions according to interfacial area generated per unit gas injection power across the operating envelope of a fixed geometry.
physics.flu-dyn
Aft sections of military cargo aircraft employ flat surfaces at high upsweep angles to accommodate ramp doors, producing flow features that affect cargo-drop accuracy, paratrooper safety, and aerodynamic performance. Fundamental studies have primarily examined near incompressible flow over a canonical surrogate consisting of a freestream aligned cylinder with a planar, sharp edged upswept base. The flow exhibits peripheral separation, a horseshoe vortex, and a counter-rotating streamwise vortex pair that persists downstream. The present investigation delineates the effects of compressibility on the wake and examines how these effects depend on basal upsweep angle. Wall-resolved large-eddy simulations are performed at Mach numbers of , , and for upsweep angles of and at a nominal Reynolds number of . For the afterbody, increasing Mach number enlarges the upstream recirculation region and delays vortex-pair formation, while these effects diminish downstream. For the afterbody, similar recirculation-region growth triggers a bifurcation at Mach~0.5 from the vortex-pair state to a broad separated turbulent wake. A descending-Mach sequence to 0.3 and 0.1 reveals hysteresis, with the separated-wake state persisting at lower Mach numbers and remaining robust to Reynolds-number variation. Thus, both states can occur at identical Mach and Reynolds numbers, with topology and pressure loading governed by Mach number history.
physics.flu-dyn
Deep learning surrogates for RANS flow prediction around airfoils face two persistent bottlenecks. A single neural architecture cannot simultaneously resolve sharp near-wall boundary layers and smooth far-field potential flow. Additionally, force prediction is undermined by the numerical instability of computing wall-normal velocity gradients from continuous-field approximations. Both of these points are addressed with a DD-RNO (domain-decomposed routed neural operator), combining a spectral geometry encoder with two physics-guided innovations: (a) a differentiable domain routing mechanism that partitions the flow field into inviscid, boundary-layer, and wake regimes—dispatching query points to specialized regional decoders, and (b) learned canonical quadrature (LCQ), which replaces unstable pressure integration with flow-conditioned, learned integration weights that predict lift and drag directly from surface pressure. On the AirfRANS benchmark, DD-RNO cuts velocity field mean-square error (MSE) by 17 () and 12 () over the strongest baseline, widening to 23 under out-of-distribution Reynolds extrapolation—evidence that the routing mechanism generalizes with the physics it encodes rather than merely fitting the training distribution. LCQ reduces drag MSE by 7.5 relative to conventional pressure integration and raises drag rank correlation from to . Ablations confirm that both components are indispensable to performance: removing domain routing increases velocity error by 8.2, and removing LCQ increases relative drag error more than 40-fold. At ~144 ms per sample—a 10,000 speedup over conventional RANS solvers—DD-RNO offers a surrogate accurate and fast enough for real-time aerodynamic design and optimization loops.
cs.LG
Accurate modeling and forecasting of power transformer thermal behavior are critical for reliability, asset lifetime, and optimized power system operation. Numerical approaches such as finite element methods (FEM) and computational fluid dynamics (CFD) offer high fidelity but are computationally expensive, require complex mesh generation, and are often impractical for real-time or large-scale applications, particularly when transformer geometries are unknown. Lumped-parameter thermal models are more practical but depend on transformer-specific thermal constants and may fail to capture dynamic responses under varying operating and environmental conditions. Purely data-driven machine learning methods, including artificial neural networks, convolutional neural networks, and long short-term memory (LSTM) networks, have shown success in forecasting transformer temperatures but typically require large volumes of high-quality training data and may produce physically inconsistent or uninterpretable results. This paper develops a physics-aware Neural Ordinary Differential Equation (Neural ODE) framework for forecasting transformer thermal behavior from real-world time-series data. Neural ODEs model system dynamics in continuous time, providing smooth trajectory prediction and a natural representation of continuously evolving thermal dynamics. A key contribution is the integration of simplified heat-transfer equations directly into the Neural ODE formulation. The model is evaluated across datasets from fifteen transformers in different regions of Norway with varying designs and cooling mechanisms. The results demonstrate that the developed Neural ODE framework provides a standardized, physics-aware, and robust forecasting approach for heterogeneous transformer units.
stat.ML
Clinically actionable, patient-specific hemodynamic assessment, specifically wall shear stress, vortex structure and pressure distributions, is critical for determining risky or unfavorable evolution in Abdominal Aortic Aneurysms (AAA). While Physics-Informed Deep Operator Networks (PI-DeepONets) show promising results in complementing established 5 tools such as Computational Fluid Dynamics (CFD), a persistent architectural challenge remains for complex 3D flows. In this direction, we propose a Modified Multi-Input Multi-Output PI-DeepONets (M3PI-DeepONet) designed for predicting unsteady flows in an idealized AAA geometry. Central to our model is the Aggregated Injection strategy, where latent representations from multiple input branches are fused prior to trunk injection, allowing the coordinate basis to adapt to multiple physical constraints. To the best of our knowledge, this is the first architecture to combine the layer-wise gating mechanism with a multi-branch operator-network topology, yielding an input-adaptive trunk basis. Additionally, we integrate the 3D Navier-Stokes equations as governing physical laws, so the model is trained based on physics-informed residuals, initial and boundary conditions, and only 0.3% of the labeled internal data together with the selected branch-conditioning signals. The M3PI-DeepONet simultaneously predicts unsteady 3D flow velocity and pressure fields with an average relative L2 velocity error below 4% and pressure error around 5% while achieving a conservative retained-cycle inference speedup of approximately 36x compared to reference CFD simulations once the branch inputs used for conditioning are available. This work advances the application of deep learning in cardiovascular disease modeling, marking step toward real-time, non-invasive clinical diagnostics.
cs.LG
Full self-attention is a strong token mixer for PDE surrogates on irregular domains, but its quadratic cost limits its use on large problems. Latent-space attention methods such as PerceiverIO, Transolver, and FLARE (Fast Low-rank Attention Routing Engine) avoid that cost by routing attention among tokens through learned latents. They compress by dot-product matching of the input tokens against learned query tokens or projection weights: once trained, the same learned templates serve every input. We remove this restriction with FLARE++, a low-rank attention architecture with input-conditioned routing queries. FLARE++ uses FLARE's own encoder to map the input tokens to query tokens, which correct the learned queries. The adapted queries then determine how that same input is compressed and redistributed. This preserves FLARE's explicit low-rank factorization and linear complexity, and expresses the complete routing operation with standard scaled dot-product attention (SDPA) calls alone. We also provide a multi-GPU context-parallel implementation that shards input tokens across devices without ever gathering the full token sequence on one of them. FLARE++ reduces FLARE's error by on average across five standard PDE benchmarks, achieving the lowest errors among the efficient models compared. The gains persist on industrial-scale DrivAerML aerodynamics and on Long Range Arena, where average accuracy rises by over fixed-query FLARE. Code is available at https://github.com/vpuri3/FLARE.py.
cs.LG
Forward and inverse modeling of parametric dynamical systems requires surrogate models that are not only accurate for state prediction, but also informative for parameter calibration. However, a systematic end-to-end differentiable formulation for coupling deep-learning-based reduced-order surrogates with variational parameter estimation remains underdeveloped. In this work, we introduce a physics-aware neural-network-based latent-space framework for reduced-order forward modeling and variational parameter estimation. The proposed autoencoder-based approach yields a differentiable surrogate that maps physical parameters to predicted flow fields through a latent representation. The observable supervision is used during offline training to encourage the latent variables to retain information correlated with system parameters, while the online inverse problem is solved in the parameter space through the surrogate-induced observation operator. The method is evaluated on two computational-fluid-dynamics benchmarks. The results show that reconstruction accuracy alone is insufficient for inverse modeling, owing to the lack of end-to-end differentiability or physics awareness for variational parameter calibration. Quantitative latent-space analysis further shows that observable supervision improves case-level separability and temporal organization of latent representations. Experiments with realistic measurement settings, including noisy, low-resolution, randomly masked, and block-wise partial observations, demonstrate the robustness of the proposed framework and show that it generally reduces calibration error and variability compared with the standard surrogate models.
math.OC
Powered landing with variable mass, free final time, and quadratic aerodynamic drag requires the repeated solution of local convex subproblems, whose main online cost lies in the long dynamics-equality chain and the inner iterations. This paper develops a condensed sequential convex approximation designed for low latency. Exact block elimination removes 217 intermediate-state components and 210 interval equations from a 31-node model, leaving 100 primal variables coupled by six terminal equalities. A low-weight energy term and fixed quadratic proximal regularization make the ideal surrogate strongly convex with predictable curvature. The inner solver is an extrapolated proportional–integral projected gradient (xPIPG) implemented with fixed-size arrays, interval solves, and a fused one-pass node map. The one-pass map is a deliberate low-cost approximation, not the exact joint proximal operator. We therefore evaluate the timed code by nonlinear trajectory residuals and independent physical checks rather than by a claim of exact KKT convergence. The single-precision C implementation completes one plan in four outer updates and 336 xPIPG updates. On an Intel Core i7-10875H, the median end-to-end solve time is \SI{374}{\micro\second} and the P99 value is \SI{512}{\micro\second}. All 100 common initial-state perturbations pass validation, and the median remains below \SI{0.7}{\milli\second} for 15–51 nodes. An independent high-accuracy first-order-hold integration gives a terminal position error of \SI{0.183}{\meter}. Within the stated model, hardware, stopping rule, and timing boundary, this is, to the authors' knowledge, the first submillisecond end-to-end sequential-convex solve for a single powered-landing trajectory.
cs.RO
Direct teleoperation of vehicles faces critical technical bottlenecks: communication latency and the operator's inability to physically perceive unmodeled environmental disturbances (e.g., aerodynamic drag, bank angles) coupled with highly nonlinear tire-road dynamics. To address these challenges, we propose a tailored unilateral teleoperation framework. The system integrates the Wave Variable (WV) approach to passively guarantee stability under stochastic delays, and an adaptive Radial Basis Function Network (RBFN) to actively compensate for vehicle-specific uncertainties. Unlike existing WV-neural network architectures designed for bilateral robotic arms, our framework features decoupled adaptive laws specifically designed for vehicle longitudinal and lateral dynamics. Furthermore, compared to model-heavy predictive controllers, the model-free RBFN offers rapid online adaptation without heavy computational overhead. Building upon our preliminary theoretical formulation, this brief paper presents comprehensive comparative analyses and real-world hardware validations. Simulation benchmarks against PID, LQR, MPC, and NMPC demonstrate that the RBFN achieves superior robustness against unmodeled disturbances while requiring orders of magnitude less execution time than MPC and NMPC, making it ideal for resource-constrained vehicle edge computing. Finally, hardware-in-the-loop experiments using a 1/10th scale vehicle over a 4G network validate the system's practical feasibility, safety, and robust trajectory tracking under physical road uncertainties.
cs.CE
Solving PDE-governed physical problems is computationally expensive, limiting the availability of high-fidelity (HF) data for training neural operators, which typically require large datasets. Multi-fidelity learning addresses this by combining abundant low-fidelity (LF) data with limited HF samples. To evaluate several multi-fidelity neural operator strategies, including two-step, residual, intermediate, and transfer learning, we conduct a systematic assessment across four PDE test cases: two standard grid-based problems, a modified Darcy flow with LF-HF governing equation discrepancies, and an unsteady smoke inflow with temporal dynamics. Most prior studies focus on grid-based LF-HF differences, but real-world discrepancies, such as those between RANS and LES models in CFD, can be more complex. To capture these more challenging scenarios, we introduce two new test cases specifically designed to emulate nontrivial LF-HF mismatch beyond simple grid-resolution differences, enabling a more realistic assessment of multi-fidelity strategies under substantial model discrepancies. We find that methods directly feeding LF predictions into HF models are prone to error amplification under large LF-HF discrepancies, particularly in autoregressive settings. In contrast, transfer learning, which uses LF-trained weights to initialise the HF model, consistently improves accuracy by providing a robust prior refined with HF data. These results highlight the importance of establishing transfer learning as a reliable approach for efficient multi-fidelity PDE surrogate modelling.
physics.flu-dyn
Neural surrogates offer a promising route to accelerating computationally expensive simulations governed by partial differential equations across science and industry. Their practical deployment, however, is limited by unreliable predictions under out-of-distribution (OOD) conditions. We develop a solver-coupled surrogate-Newton framework that uses surrogate predictions as high-quality initial guesses for Newton-Krylov iterations, thereby combining rapid global flow-field prediction with high-accuracy numerical convergence at the terminal stage. On an OOD benchmark comprising geometries sampled from actual transonic airfoil optimization trajectories, the framework lowers the median residual L_2 ratio by over seven orders of magnitude while substantially reducing field and aerodynamic errors. In practical supercritical airfoil optimization, it improves online prediction reliability while achieving a 15.5-fold generation-level speedup over CFD. We further test the framework's extension to three dimensions using a flying-wing dataset. Together, these studies demonstrate the potential of surrogate-Newton coupling to deliver accurate, efficient and scalable steady CFD across industrial workflows.
stat.AP
In early 2022, FIA mandated the change to the technical regulations for Formula 1, requiring to design cars around ground effect aerodynamics to mitigate the performance loss incurred by driving in the turbulent "Dirty Air" of another car. The reduction of the penalty was subsequently claimed, but never formally evaluated on a per lap basis. We gathered 100486 racing laps for the years 2021-2025 using the FastF1 API, built a compound-specific physics correction for Dirty Air to account for fuel and tyre wear, and evaluated the change introduced by the 2022 regulation using both OLS regression with clustered standard errors and Causal Forest Double/Debiased Machine Learning. We repeated the analysis using season as a categorical variable, applying a generalized circuit filter to account for a changing calendar. Both methods produced similar results, demonstrating no statistically significant change to the Dirty Air penalty when combining 2022-2025 data together (OLS interaction term beta = -0.0025, p = 0.803; Causal Forest mean effects of 0.019s and 0.018s per second of following distance in 2021 and 2022-2025). We tested the hypothesis, that the penalty reduction may have been concentrated in 2022 and partially eroded through 2023-2025. We found no season for which the change was different from the one observed in 2021 in the OLS analysis, while several seasons were significantly different from zero in the Causal Forest analysis, in a non-monotonic pattern, not following the theorized erosion timeline. We conclude there was no statistically significant change to the Dirty Air penalty under the new regulation, both in pooled and season-specific analysis. By evaluating the real-world racing results, we demonstrate that even if an aerodynamic advantage exists, it did not translate into a measurable improvement over following distance, where other factors can swamp such an effect.
physics.space-ph
Thermospheric mass density governs aerodynamic drag in low Earth orbit and is a primary source of uncertainty in orbit prediction and conjunction assessment, particularly during geomagnetic disturbances. We present AETHER-P3 (Accelerometer-driven Estimation of THERmospheric density-A Physics-Informed Probabilistic Prediction Platform), a machine-learning-based global thermospheric density forecasting model that provides multi-step forecasts up to 6 hr ahead using a 3-hr input window, with predictive uncertainty estimates. AETHER-P3 formulates thermospheric density forecasting as a sequence-to-sequence regression task conditioned on recent space weather evolution and a user-specified sequence of future times and locations. To enhance physical consistency and generalization, AETHER-P3 incorporates JB2008 and NRLMSISE-00 density estimates evaluated at future locations, along with solar, geomagnetic, and solar-wind drivers. The network employs dual recurrent encoders and an evidential Normal-Gamma output head to jointly estimate forecast mean and uncertainty. The model is evaluated using independent satellite test cases spanning quiet, moderate, and extreme geomagnetic conditions. During quiet periods, AETHER-P3 achieves high forecast skill (R=0.95). Under moderate activity, strong skill is retained (R=0.93), with reduced physical-domain errors than empirical baseline models. During extreme storm conditions, deterministic forecast skill degrades as expected yet remains robust (R=0.89-0.90). Predictive uncertainty remains well calibrated across all regimes. These results establish AETHER-P3 as a practical, low-latency, uncertainty-aware capability for thermospheric density forecasting that supports orbit prediction, drag-risk assessment, and operational decision-making over its validated altitude range of approximately 300-520 km, with highest confidence in the data-rich 400-520 km region.