Computational Engineering, Finance, and Science
Covers applications of computer science to the mathematical modeling of complex systems in the fields of science, engineering, and finance. Papers here are interdisciplinary and applications-oriented, focusing on techniques and tools that enable challenging computational simulations to be performed, for which the use of supercomputers or distributed computing platforms is often required. Includes material in ACM Subject Classes J.2, J.3, and J.4 (economics).
sort pith recommended most recent
One inference-time knob trades expected gains for run-to-run steadiness in EV cooling-channel design, confirmed by CFD.
· “Risk-Aware Generative Inpainting for Optimized Design Editing of EV Battery Cooling Channels”
Thin hinge design sends stiffness negative, spreading impact into vibration that leaves a dropped egg intact.
Single model returns displacement fields and frequencies 1,000x faster than finite-element analysis.
· “Learning Metamaterial Eigenmodes with Wavelet-Encoded Fourier Neural Operators”
A three-level risk quadrangle framework lets practitioners encode aversion to rare high-consequence outcomes, with sensor placements that…
· “Risk-Aware Goal-Oriented Bayesian Optimal Experimental Design”
A process-mining agenda that mines logs into evidence packages with the authority to act, defer, ask, or refuse.
· “From Event Logs to Governed Action: A BlueSky Agenda for Agentic Process Mining”
Friction-aware yield with hardening predicts printed-wall collapse to within one layer.
In an LLM-agent simulation, wallet-level flags leave the cross-agent communication and later exit alive.
· “Your Agent Says Yes: Interpreting Adversarial Market Behavior Beyond Individual Transactions”
No per-scan optimization: model trained once generaliess across MRI and CT datasets with state-of-the-art fidelity
A comprehensive review of weak boundary condition methods derived from Nitsche's method for advection-diffusion and Navier-Stokes…
· “A review of weakly enforced Dirichlet boundary conditions in computational flow analysis”
Sliders set leverage, render bounds route picks, and mechanical exits beat every discretionary policy; no fleet shows an edge.
A multi-agent LLM system outperforms single agents on 46-part chamfer distance benchmark by compounding visually checked transitions.
Elliptical layer models match real collapse height within one layer; rectangles can mislead by tens of percent.
A differentiable hybrid model for crystal dissolution discovers physically consistent kinetics and optimizes temperature profiles in a…
Spectral dynamic block activation skips unnecessary updates while keeping accuracy, tested on three scientific applications.
· “sp-DBA: a general framework for adaptive transform-domain computation”
Execution feedback and tutorial knowledge drive performance; role decomposition and scripted reflection add nothing.
· “What Do CAE Simulation Agents Really Need Beyond a Generic Harness?”
FinRAG-QA tests 999 cross-bank questions on 10 indicators; optimized pipeline beats baselines by over 30 percentage points.
· “Enhancing Financial Question Answering: A Novel Benchmark Dataset of Banks' financial statements”
Unsupervised framework uses only reaction forces and surface displacements, bridging neural inelasticity and full-field experiments.
· “Calibration of neural viscoelastic models via full-field data”
A three-dimensional framework reveals hidden role differences among currencies, gold, and Bitcoin.
State-dependent weights shrink nonlinear finite element models to near the theoretical minimum of sampled points.
Trained on 5-fin chips, it extrapolates to 38 fins and screens a device in seconds, no retraining.
· “Mesh-Native Physics-Informed Graph Surrogates for TCAD-in-the-Loop Design Space Exploration”
GP-based preference elicitation combined with multi-objective RL captures how European and Texas investors trade off returns against…
· “Eliciting ESG Preferences for Reinforcement Learning-Based Portfolio Optimization”
A PINN framework for depth-averaged granular avalanche flow on curved topography is validated against experimental data, showing that…
Combines four particle methods to simulate reaction-diffusion and shape change on deforming surfaces in fluid
· “A meshfree solver for coupled bulk-surface problems with self-organizing surface geometry”
Trained on 50-step paths, the operator holds about 2% error at 1,000 steps while step-wise surrogates drift to 10-14%.
Physics-enhanced framework visualizes degradation functions, boosting trust for predictive maintenance in rotating machinery.
One conditional GFlowNet serves the whole risk-composed reward family at 0.37-2.7x a perfect sampler's error floor.
· “Risk-Sensitive Reward Composition for Conditional GFlowNets”
Across 17,000 coins, timing—not wording—separates coordinated pumps from organic reactions
· “Don't You Know, Pump it Up! Investigating Cryptocurrency Manipulation in Telegram-Driven Activity”
Context-aware vulnerability ranking and a two-stage detector cut false positives to 1.2% for small firms.
Latent-space diffusion also bends 3D turbine blades to new constraints without retraining.
· “Aerodynamic Shape Design Space Exploration with Deep Latent Diffusion Model”
HarmoCore places a generative prior in a compact Functional Tucker core, enabling accurate 2D and 3D reconstruction from sensor densities…
· “HarmoCore: Functional Latent Diffusion for Sparse Reconstruction of Oscillatory Wave Fields”
Automatic canonicalization and KKT reformulation turn hierarchical models into tractable nonlinear programs.
MUFASA improves tumor diagnosis, subtyping, biomarker prediction, and survival modeling by excluding uninformative tiles while preserving…
Weighted similarity edges join disconnected components as documents arrive, at 92 percent fidelity for a quarter of the storage.
Only one of two pretrained models transfers to stroke; zero-shot fails completely
The paper establishes that the combination of surface DC-PSE and EDAC provides a consistent, convergent, high-order numerical solver for…
Pairing each confidence map with its own fitted scale improves enemble returns by 9.2 pp/yr; cross-applying degrades performance.
· “Ordinal Gates, Cardinal Bets: Matching LLM Confidence to the Financial Decision Operator”
Plan–Execute–Interpret architecture eliminates two failure classes by construction, with scoped ontology facts correcting LLM…
· “Domain-Grounded Tool Orchestration for LLM-Guided Scientific Analysis”
Derivation shows SHoSP weight is a material-mass perturbation budget, explaining stress and hinge suppression in topology optimization.
A maximum-likelihood unit lets stablecoins track real value, not fiat, and cuts turnover fourfold in backtests.
· “Redefining Stablecoins from Nominal to Real Value: A Maximum Likelihood Approach”
Validator rejects factors whose price evidence contradicts their hypothesis; the survivors beat all baselines out of sample.
· “FaVOR: LLM-Based Agentic Framework for Factor Mining via Empirical Validation”
Hybrid physics–AI digital twin synchronizes at 1 Hz and forecasts power over full operating cycles.
Parameter-space method is more accurate, efficient, and extends to higher dimensions.
· “Intrinsic Finite Element Methods for Fluids on Riemannian Manifolds Compared with Surface FEM”
PAVA converts continuous cyclone fields into discrete, checkable labels, boosting multimodal accuracy from 45% to 67%.
· “OmniClimate-TC: Physics-Aware Visual Abstractions for Multimedia Reasoning over Tropical Cyclones”
A calibration-based reconciliation method for reduced-state scientific twins that uses future boundary data to correct the provisional…
· “QoI-Aware Provisional Rollout and Retrospective Reconciliation for Reduced-State Scientific Twins”
Matrix product states compress 1024³ DNS fields while preserving energy, dissipation, and structure functions within a few percent.
· “A priori Assessment of Tensor-Network Encoding for Isotropic Turbulent Flows”
The same price stream can look random to one trader and fully predictable to another, and the paper formalizes why.
· “Pragmatic Information, Computation, and the Efficient Market Hypothesis”
Cholesky-constrained network predicts stiffness and density, enabling efficient two-stage topology optimization of graded superimposed…
A combined method takes first place on the ConStellaration geometric task in May 2026.
· “Towards Stellarator Geometry Optimisation for Nuclear Fusion”
DG-SWEM with a simple parametric rain model catches 25 of 104 Beryl high-water marks, vs 2 with surge alone.
The 2D model's sensitivity shapes the kernel, cutting costly 3D RANS solves and making online inference instant.
· “Dimension Bridging for 3D RANS with Neural Network Accelerated Gaussian Functional Regression”
Adaptive resampling and variance-reduced guidance lift reward on both benchmarks while held-out structure checks hold.
Community workshop distills four themes into eight actions for trustworthy, sustainable, and resilient ecosystems.
φ-DeepONet learns mappings with discontinuities in inputs and outputs by combining multiple branch networks with a nonlinear interface…
Kronecker-factored transitions cut complexity from exponential to linear while keeping global dynamic programming solutions intact.
It learns to roll out finite element states autoregressively for faster manufacturability checks in sheet metal design.
EqOD first tests for Galilean invariance on the data, drops impossible terms, then runs stability selection, outperforming standard sparse-P
· “EqOD: Symmetry-Informed Stability Selection for PDE Identification”
Kokkos code matches specialized single-GPU speed while enabling billion-particle spectral plasma runs on three supercomputers
· “A Performance-Portable, Massively Parallel Distributed Nonuniform FFT”
FeatureFox combines binary edge classification on B-Rep graphs with connected-component instance recovery to deliver sample-efficient…
Adjusting periodic boundaries for thickness-direction surfaces yields stable effective stiffness matrices
· “A 99-Line Homogenization Code for Lattice-skin Plate Structures”
Nesterov-accelerated projection with subspace embeddings preserves accuracy while cutting full-model runtime from over a minute to under 3 s
Fine-tuned models inside a feedback loop outperform larger general models on verified FEniCS scripts for solids, fluids, and multiphysics.
· “ALL-FEM: Agentic Large Language models Fine-tuned for Finite Element Methods”
Closed-form projection replaces iterative fitting and unifies logarithmic, Saaty, and SVD windowing.
Trained on synthetic workloads, it migrates 7.7x fewer slices than reactive balancing at similar runtimes.
· “Scalable Multi-GPU Simulation of 3D Multicellular Growth with RNN-Based Workload Balancing”
Default sampling hides a confounder: a standardized three-step recipe lifts a weak baseline past specialist models.
· “Unlocking Multimodal Protein Language Models at Inference Time”
A co-rotational formulation with an auxiliary straight reference handles arbitrary initial shapes under large rotations.
On a 33-bus feeder, fairness raises curtailment from 2.11 to 5.72 MWh while holding the max cumulative ratio to 11.2%.
Dual-horizon equations fall out of the action principle; asynchronous stepping cuts force evaluations by 30%.
A Gaussian-process derivative match seeds a multi-shooting refinement that enforces global dynamical consistency.
· “Gaussian process learning with flow map refinement for parameter estimation in dynamical systems”
A trainable proxy picks which evidence to fetch at each reasoning step, beating prior self-evolving agents on BioHopR.
Gemini, Claude, and Grok passed an AI-generated geothermal benchmark.
· “Decision-Support and Modeling with Large Language Models for Geothermal Well Arrays”
Budget-aware recipe search beats literature defaults on DNABERT2 and Nucleotide Transformer benchmarks.
· “GenomeHarness: Harnessing Al Agents for Reliable Adaptation of Genome Language Models”
Frozen features stay near chance; updating the visual encoder makes predictions depend on intact scene layout.
RWR interleaves global latent attention with local message passing, winning on flow fields, gradients, and wall shear
· “Read, Write, Relax: Why Neural PDE Surrogates Need Both Global and Local Processing”
Prompted with benchmark data, an LLM reproduces SmartSolve's LU algorithm-selection tree.
· “Data-Driven Dynamic Algorithm Dispatch with Large Language Models”
The model works with sparse wellhead data; the same tool shows why single-well geothermal repurposing doesn't pay back.
Two stabilized integral-equation methods fix low-frequency breakdown, so near and far fields follow the Lorenz gauge down to DC.
In an oracle upper-bound test, the model reaches Dice 0.46 versus 0.25 for standard thresholds.
· “An Imaging-Informed Reaction-Diffusion Model of Infarct Growth”
On FinanceBench, ontology-grounded retrieval lifts citation traceability while answer correctness stays level with BM25.
POT3D scales across multi-GPU nodes using only standard 'do concurrent' loops and unified memory.
1,584 tokens and 27 daily variables show synchronized market-wide peaks and point to ecosystem signals for forecasting.
· “Time Series Forecasting based on Solana Digital Asset Dataset”
