Over the last decade, widely publicized AI breakthroughs have disrupted various areas of science and technology. Many of the most credible and transformative advances have followed a “common task framework” (CTF) paradigm, in which a clearly specified machine learning task is shared with a research community through standardized datasets, baseline algorithms, and evaluation metrics. Within this framework, researchers compete to improve performance, enabling rapid identification and adoption of superior methods. Despite its success, the CTF paradigm is not yet universal. Some fields lack standardized benchmarks, shared code and data, or consistent evaluation practices, leading to fragmented progress and difficulty comparing claims. Fields that have embraced CTFs repeatedly demonstrate accelerated and verifiable improvement. We argue that dynamic systems research is ready to adopt this approach. We introduce CTF4Science, a framework for defining tasks, publishing datasets and benchmarks, and tracking performance across prediction, denoising, and parameter extrapolation problems.
Artificial Intelligence and Machine Learning have become central tools in scientific and engineering computation. Their adoption has accelerated across physics, materials science, structural engineering, thermal modeling, and control systems. Despite these benefits, the use of Artificial Intelligence introduces nondeterministic behavior that conflicts with longstanding expectations of reproducibility, traceability, and predictable numerical performance. Regulatory frameworks in medicine, aerospace, automotive safety, and nuclear engineering require deterministic computation for verification and certification [1], [2], [3]. This article examines the controversies that arise when Artificial Intelligence is applied to scientific and engineering workflows. It analyzes the sources of nondeterminism in hardware, software, and models. It presents the Mediated Deterministic Architecture (MDA) as the primary engineering contribution of this work, an architecture that constrains nondeterministic components behind deterministic validation layers governed by three design principles, independence, directional sensitivity, and extrapolation awareness, and provides a precise guarantee: every output the system emits satisfies an explicit set of physical invariants, by independent validation or by certified fallback. The article evaluates the implications of these controversies and outlines future directions for responsible integration of Artificial Intelligence into high assurance computation.
We examine super-resolution models, which predict system state at high resolutions given coarse resolution information, and hold great potential for multi-scale scientific applications. We classify existing models into implicit and explicit super-resolution models, and consider their theoretical underpinnings and state-of-practice from a spectral viewpoint. For implicit (zero-shot) models we identify a fallacy in their claim of generality; they require data or prior knowledge of interactions between scales for predictions to be unique. Explicit models, such as those used for turbulence, have issues with spectral bias, affecting predictions of fine-scale energy. While a few recent works have proposed remedies for spectral bias, we contend that capturing total energy is a necessary, but not sufficient condition, for accuracy. To accurately describe fine scales, models must also preserve local isotropy and physics invariants.
Motivated by the ongoing debate over whether physics-informed neural networks (PINNs) can serve as universal solvers for partial differential equations (PDEs), this paper uses elliptic PDEs as a concrete setting to describe a family of physics-preserving neural network (P2NN) methods. Beyond being physics-informed, P2NN methods are designed to preserve the underlying physical principles which are often violated by PINNs. Specifically, P2NN methods employ neural networks as approximating functions, an equivalent formulation of the governing PDEs, and physics-preserving numerical differentiation operators. For diffusion-reaction equations, the Ritz and the dual formulations, associated with the primal and dual variables respectively, provide natural optimization principles and hence preserve physics. For convection-diffusion-reaction equations, we examine the Bramble-Schatz and firstorder system least-squares (LS) formulations and assess their applicability. Finally, we emphasize that a viable NN-based method also hinges on accurate numerical integration and physics-preserving numerical differentiation. These components are essential, nontrivial, and are discussed within the proposed P2NN framework.
This study presents a review of Mapper in applications for patient stratification. We identify successes and shortcomings of Mapper applications and connect these to problems in ML for science more broadly.
Foundation models (FMs) have achieved substantial success in generalizing across tasks without problemspecific training or fine-tuning. However, many critical applications in mechanics and computational science require not only accurate predictions but also reliable uncertainty quantification (UQ). Herein we investigate the UQ capabilities of tabular FMs in regression tasks through a comprehensive empirical study comparing Tabular Prior-Data Fitted Networks (TabPFN) against Gaussian processes (GPs). We systematically evaluate these two methods across a host of regression problems with varying complexity, dataset sizes, and input dimensionalities. We use a default setting to build all the GPs and for a fair comparison against TabPFN v2.5. Our findings highlight an important trade-off between explicit and learned priors: while TabPFN achieves highly competitive performance for complex, high-dimensional problems with sufficient data, GPs often provide superior predictive accuracy and UQ in data-scarce settings. Moreover, when the chosen kernel constitutes a good prior for the underlying function, GP performance can substantially exceed that of TabPFN. Our results can be reproduced from https://github.com/kianswarehouse/GPvsPFN.
Physics foundation models are evaluated primarily on in-distribution accuracy at training resolutions. This measures in-domain approximation and not whether models have learned continuous physics. We propose validation protocols along two axes: out-of-distribution generalization (do models trained on easy problems generalize gracefully to hard ones?) and temporal consistency (do models produce coherent dynamics under resolution changes?). These benchmarks are straightforward to implement and would clarify whether current models are learning physics or memorizing discrete computational artifacts.
Wildfire management is a persistent, time-critical, and data-intensive challenge that includes emergency response, strategic planning, and long-term landscape management. Urgent computing has traditionally emphasized rapid allocation of scalable computing resources during disasters, but speed alone is insufficient. In wildfire contexts, readiness for sustained integration of science-driven tools with operational decision making is equally critical. This article reframes urgent computing for wildland fire as a readiness problem enabled by composable cyberinfrastructure that can be continuously exercised, adapted, and reused across the full wildfire lifecycle. It argues that enabling platforms, powered by urgent computing, are essential for such readiness. Drawing on the Wildfire Science and Technology Commons, the article presents three contrasting examples: Firemap for real-time fire response, BurnPro3D for prescribed fire planning, and Points2Pano for continuous vegetation analysis. These examples demonstrate how composability across data, models, workflows, and compute resources supports both emergency response and mission-critical planning, enabling operational readiness.
The term foundation model is rapidly entering scientific machine learning, yet its meaning in computational science remains inconsistent. Models are increasingly labeled foundational without evidence of broad generalization across physical systems, geometries, boundary/initial conditions, tasks, or operating regimes. We propose a definition of foundation models tailored to computational science and identify core and desirable characteristics, drawing inspiration from classical foundational methods (FEM/FVM/FDM) that achieve wide reuse through problem-independent structure and principled enforcement of governing equations. We argue that scientific foundation models face unique barriers relative to language and vision models, including field-valued data, mesh and geometry dependence, heterogeneous physics, expensive adaptation, limited benchmarks, and strict requirements for interpretability and physical consistency. To make claims measurable, we introduce a reviewer- and author-friendly checklist for reporting training breadth, generalization axes, adaptation cost, extrapolation protocols, and scientific consistency checks. We also highlight a constructive hybrid path: local learning with global physics enforcement.
Research software has been categorized for various goals. One fundamental dimension of such categorizations is the role that the software plays in the research process. Recently, a new role category has emerged: technology research software, which covers research software developed in technology research. Until now, this category of technology research software has often been overlooked and neglected within the research software engineering community. In this article, we explain technology research software and its primary subroles. Technology readiness levels are an established method of estimating the maturity of technologies, including software systems. For technology research software, these readiness levels define secondary subroles. To illustrate the concept of technology research software and to make it more tangible, we present examples of research software that, depending on its specific use within or outside of research, take on the role of technology research software as well as that of another research software category.
Hydraulic erosion models simulate the gradual wearing away of terrain by flowing water, a phenomenon critical to realistic landscape generation and geomorphological research. Among the various computational approaches, particle-based methods provide a simple yet powerful framework for modeling erosion dynamics on discrete heightmaps. In this article, we implement a two-dimensional particle-based hydraulic erosion algorithm that models the transport of sediment using simulated water droplets that traverse heightmap data. This process demonstrates how local particle-terrain interactions lead to large-scale landscape features such as valleys, gullies, and mountain ridges.
Large language model-based tools are rapidly becoming part of everyday scientific programming. Here, we discuss converging findings from surveys and focus groups of scientific programmers about usage trends, tool choices, and emerging concerns. We argue that rising adoption will strain existing infrastructure for ensuring the rigor and scientific validity of scientific software, and discuss approaches to meeting this urgent challenge.
Data-intensive science and artificial intelligence-enabled workflows now shape research and teaching, yet faculty at primarily undergraduate institutions (PUIs) face tight resources and steep barriers to adopting advanced computing tools. This article reports stewardship lessons drawn from interviews with three PUI faculty who use two National Science Foundation-supported cyberinfrastructure ecosystems: Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support and the National Science Data Fabric, which are relied on to support teaching and research. Across fields spanning physics, computer science, and environmental science, the faculty converged on three practical patterns: 1) lower the barrier to first success with low-friction onboarding and ready-to-run examples, 2) make support human and immediate through visible help channels and rapid triage, and 3) make science engaging by bringing real datasets and problems into student projects. We conclude with a compact stewardship checklist for teams building and operating national-scale computing and data services to make computing practical for everyone.
Research communities are often relatively small and somewhat insular. This is not to say that the people are unfriendly to outsiders or newcomers, but more a reflection that people like to spend some time with long-term friends they see only a few times a year. Finding a way to encourage new community members to interact with more senior, long-term members requires a catalyst. Using something to equalize status and encourage people to talk can break down barriers and enable easier connection building and ultimately growing an integrated community rather than one split across generational lines.
Scientific machine learning (SciML) often operates in ill-conditioned, weakly identifiable regimes due to limited data or indirect observations. In such settings, optimization and inference are highly sensitive to the starting point, making initialization–often under-reported–a consequential degree of freedom. Random initialization is not a neutral default as it induces an implicit prior over candidate solutions and can systematically bias the result, producing large run-to-run variability. We formalize this view by treating initialization as a hidden confounder in SciML and develop a unifying theory for structure-aware initialization via numerical continuation, constructing warm starts from related problem instances. Across representative tasks, including physics-informed neural networks (PINNs), maximum likelihood estimation (MLE), and variational inference (VI), warm starts have been shown to consistently reduce optimization effort and improve reliability.
High-consequence decision making demands peak performance from individuals in positions of responsibility. Such executive authority bears the obligation to act despite uncertainty, limited resources, time constraints, and accountability risks. Tools and strategies to motivate confidence and foster risk tolerance must confront informational noise and can provide qualified accountability. Machine intelligence augments human cognition and perception to improve situational awareness, decision framing, flexibility, and coherence through agentic stewardship of contextual metadata. We examine systemic and behavioral factors crucial to address in scenarios encumbered by complexity, uncertainty, and urgency.
Artificial intelligence (AI) and machine learning (ML) models are increasingly being discussed, both explicitly and implicitly, as potential replacements for traditional scientific computing methods, raising fundamental questions regarding their suitability for scientific computation. This article critically examines these claims by assessing whether contemporary AI models satisfy the practical requirements of scientific computation. The analysis shows that while AI models offer substantial value as computational accelerators and surrogate components, they lack the systematic error control, generalization guarantees, and physical accountability required for reliable replacement. This gap explains why replacement narratives remain controversial and motivates a more precise and responsibility-aware role for AI in scientific and engineering workflows.
Convolutional Neural Networks (CNNs) have become the most common approach for automatic image classification. However, their complex and unintuitive nature makes them vulnerable to biases that are very difficult to pinpoint, and can mislead even experienced researchers. Here we study and profile possible reasons leading to such CNN bias. Experiments are performed to mimic the generation of image datasets, leading to high classification accuracy even when no information about the objects is present in these images. The analysis shows that even when using the exact same conditions for all images, CNN architectures can still provide classification accuracy far higher than mere chance. In fact, these biases are common in many of the commonly used benchmark datasets. The bias depends on the order of imaging, the use of camera settings, the room temperature, and the CNN architecture. The datasets created for the experiments are publicly available.
In human space exploration, complexity of vehicles compounded by hostile elements of the space environment can turn otherwise benign anomalies into outright emergencies. On missions in low Earth orbit, such anomalies are routinely resolved by flight controllers at the mission control center on the ground. The rigorous data-driven decision-making process that mission control engages arguably epitomizes the essence of urgent science. With destinations of expeditions reaching beyond low Earth orbit, light time-delayed communication will render Earth-based support more difficult and necessitate technological solutions backed by urgent computing onboard future vehicles. This article presents case studies to illustrate the type of problems often encountered in human space exploration and how human-centered urgent computing will be needed to help astronauts manage unanticipated safety-critical situations that have so far been handled in real time by the large team of experts on the ground.
Physics-informed neural networks (PINNs) encode physical models such as partial or ordinary differential equations into the loss functions of neural networks. Over the past several years, PINNs have seen applications in solid mechanics, computational fluid dynamics, material modeling, and dynamics and control, covering a range of disciplines in science and engineering. However, PINNs, at least in their naïve form, generally struggle to converge to ground truth solutions, fail to strictly enforce known physical laws, and suffer from various other numerical issues (including those inherited from learning pipelines and algorithms). Thus, we have found the community to be divided over the application of PINNs to real-world science and engineering problems. This article addresses this controversy by briefly reviewing the promise of PINNs, by examining various failure modes of PINN-based pipelines, by deriving and applying PINNs for power system simulation as a case study, and finally, by suggesting future directions for research in this area.
