PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Intelligence in Mechanical Engineering: Foundations, Opportunities, Challenges, and Responsibilities

A. Cunha Jr, C. Nataraj, L. M. Lorenzo, H. Panetto, A. Scotti, J. V. C. Vargas

arXiv:2610.33105v1cs.AIcs.CEcs.LGstat.AP

TL;DR

Mechanical engineering can use AI across research and practice, but responsible integration must account for computation’s energy demands and the need for physical consistency and human expertise. The article synthesizes applications and constraints into principles for responsible adoption, concluding that AI should augment engineering capability while remaining physically grounded, validated, and energy-aware.

  • Problem

    AI offers opportunities across mechanical engineering, but responsible integration must address energy demands and the continued need for physical knowledge, validation, and human expertise.

    Source: §1 · p. 2, §7.7 · p. 15

  • Method

    The article synthesizes engineering applications and constraints into principles and recommendations for responsible AI use across practice, curricula, and research.

    Source: §12 · p. 20

  • Results

    The paper concludes that reliable engineering AI requires physical constraints and hybrid learning, while responsible deployment must treat energy awareness as a design variable.

    Source: §12 · p. 20

  • Takeaways & Limitations

    Engineering AI should augment human capability through physical consistency, validation, uncertainty awareness, energy stewardship, and professional oversight.

    Source: §12 · p. 20

    Agentic systems remain dependent on human expertise, scientific principles, validation, physical knowledge, and professional oversight.

    Source: §7.7 · p. 15

Abstract

from arXiv · show

Artificial intelligence (AI) is reshaping mechanical engineering (ME), from modeling, simulation, and optimization to diagnostics, design, and technical communication. Drawing on a round-table at the 28th ABCM International Congress on Mechanical Engineering (COBEM 2025), this article presents a structured perspective for the Brazilian and international ME communities. Rather than an exhaustive survey, it examines how AI can enhance engineering practice without compromising scientific rigor. We distinguish statistical learning from human cognition and discuss the capabilities, limitations, and emerging agentic forms of AI. We then consider industrial adoption, workforce changes, and AI-assisted scientific discovery, alongside the energy and infrastructure demands of AI computation and their implications for sustainability and thermal management. The discussion emphasizes the integration of physical knowledge, structured reasoning, and domain expertise into AI systems, particularly in safety-relevant applications. Topics include physics-informed and hybrid learning, physically constrained modeling, validation, uncertainty quantification, and certification. We also examine educational, epistemological, and ethical challenges, including curriculum reform, human-agent collaboration, professional accountability, and the preservation of engineering knowledge. We conclude with principles for responsible AI adoption in ME: physical consistency, rigorous validation, explicit treatment of uncertainty, energy sustainability, human oversight, and ethical responsibility. AI should complement, not replace, engineering judgment.

1 Introduction

AI is reshaping mechanical engineering, bringing opportunities across research and professional practice alongside physical, safety, energy, and governance constraints. This article offers a practice-oriented synthesis of how AI can strengthen engineering while preserving rigor and judgment.

1 Introduction

  • AI can support work from concept generation and simulation to maintenance and technical communication, while lowering barriers to computational capabilities and contributing to productivity gains.These opportunities span advanced engineering and routine professional activities.Source: §1 · p. 2
  • In mechanical engineering, AI decisions affect physical artifacts and systems, making safety, reliability, compliance, traceability, and accountability central conditions for constructive use.Engineering integration also requires attention to model validity, life-cycle consistency, and interoperability.Source: §1 · p. 2
  • AI computation creates energy, carbon, and water demands, making energy efficiency, resource availability, and environmental impact part of responsible adoption.The paper connects these concerns to mechanical engineering expertise in energy conversion, thermal management, and efficiency.Source: §1 · p. 2
  • The article synthesizes a COBEM 2025 round-table for Brazilian and international mechanical engineering communities rather than attempting an exhaustive survey.Its stated aim is to strengthen engineering quality and judgment while mitigating risks and avoiding superficial automation optimism.Source: §1 · p. 3
  • The article establishes conceptual grounding in AI learning paradigms and limitations, distinguishing statistical learning from human cognition and introducing agentic systems for engineering workflows.It also considers their relevance to scientific discovery.Source: §1 · p. 3
  • Its technical perspectives emphasize structured knowledge, reasoning, physical consistency, and hybrid methods, including physics-informed approaches for dynamic-system modeling and diagnostics.These methods address sparse or noisy data, interpretability, robustness, and long-horizon predictive fidelity, particularly in digital twins.Source: §1 · p. 3
  • Representative applications include structural health monitoring, fatigue and fracture mechanics, additive manufacturing, computational mechanics, materials design, predictive maintenance, and agentic workflows.The discussion emphasizes integrating data-driven methods with physical modeling, simulation, uncertainty quantification, and engineering judgment.Source: §1 · p. 3
  • The article also addresses preservation of engineering knowledge, curriculum modernization, ethics, accountability, validation, uncertainty quantification, and certification.Its final recommendations span industry, university curricula, research, and human–AI collaboration.Source: §1 · pp. 3–4

2 How contemporary AI operates

Contemporary AI encompasses data-driven learning, generative models, and agentic systems that can coordinate tools and multi-step engineering workflows. Their capabilities remain bounded by task scope and unreliable reasoning, so engineering use depends on human expertise and validation.

2 How contemporary AI operates

  • Supervised, unsupervised, and reinforcement learning differ in whether training uses labeled outputs, unlabeled structure, or reward-guided interaction.The paper relates these paradigms to classification, anomaly detection, control, and sequential decision-making.Source: §2 · p. 5
  • Agentic systems combine language models with memory, planning, tool use, and feedback to coordinate multi-step workflows toward specified objectives.They can decompose objectives, invoke computational tools, evaluate intermediate results, and adapt subsequent actions.Source: §2 · p. 5
  • Agentic AI may support automated simulation campaigns, design exploration, digital twins, and laboratory experimentation, but remains dependent on human-defined objectives, domain knowledge, physical models, and validation.Potential benefits include reducing time spent on simulation campaigns, design exploration, data analysis, and model development.Source: §2 · p. 5
  • AI systems support engineering applications, but current progress on demanding evaluations does not establish robust general intelligence or reliable real-world reasoning.The article notes strong benchmark trajectories on MMMU, GPQA, and SWE-bench, alongside continued difficulties with complex planning and reasoning under uncertainty.Source: §2 · pp. 5–6
  • Most current AI systems are task-specific and perform poorly outside their intended scope, while versatile general-purpose systems remain an open research challenge.The paper notes that multimodal capabilities remain limited.Source: §2 · pp. 5–6
  • AI operations described as planning, learning, or memory are statistical and computational processes, not evidence of human-like understanding, intention, or consciousness.The paper warns that anthropomorphic interpretations can encourage misplaced trust and unrealistic expectations about reliability and accountability.Source: §2 · p. 6
  • AI systems operate within the limits of their data, architectures, and training objectives rather than reasoning or comprehending as engineers do.The distinction matters in engineering, where reliability, interpretability, accountability, and physical consistency are critical.Source: §2 · p. 6

3 Industrial adoption of AI and workforce transformation

Organizations are expanding AI adoption, but reported use remains concentrated in generative applications and its broader organizational impact is described as limited. Workforce change and the need to adapt skills and curricula accompany this early-stage transition.

3 Industrial adoption of AI and workforce transformation

  • AI use in at least one business process rose from 55% in 2023 to 78% in 2025, while generative AI use increased from 33% to 71%.Reported applications remain concentrated in text generation (63%), image generation (36%), and code creation (27%); the passage characterizes the overall impact as not yet truly transformative.Source: §3 · p. 7
  • Organizations report a gap between widespread generative-AI experimentation and limited measurable operational impact.The cited analysis describes most organizations as having piloted AI, while only a small fraction report measurable economic value.Source: §3 · p. 7
  • The article attributes limited impact largely to early adoption and anticipates broader applications as industries integrate AI with scientific models, data, and simulation tools.It frames AI as an enabler of engineering activities rather than a replacement for established methods and expertise.Source: §3 · p. 7
  • Industrial AI also operates at the edge, where latency, energy efficiency, reliability, and limited computing resources can matter more than model size.Examples include smart sensors, monitoring, condition assessment, predictive maintenance, and control.Source: §3 · p. 7
  • Changing workforce needs motivate curriculum updates that combine AI and data-science skills with creativity, collaboration, critical thinking, sustainability awareness, and ethical judgment.The passage describes these competencies as complementary to technical skills.Source: §3 · p. 7
  • The industry perspective describes adoption as early and current applications as largely focused on text, image, and code generation, while anticipating broader adoption.It identifies universities and academia as central to workforce preparation.Source: §3 · p. 7

4 Energy challenges and opportunities for AI

AI’s expanding computational demands are making energy, cooling, and infrastructure central constraints on deployment. The section surveys technical options and regional considerations, emphasizing integrated planning rather than a single remedy.

4 Energy challenges and opportunities for AI

  • High-utilization computing hardware requires reliable power and generates heat, while cooling and power-conditioning systems add to facility electricity demand.The figure traces power from primary sources through conditioning to computing hardware and thermal management.Source: §4 · p. 8, Figure 3 · p. 8
  • Approximately 4.4% of U.S. electricity demand in 2023 could rise to 12% toward decade’s end as data-center workloads grow.The projection is conditional on current trends and driven largely by AI-related workloads.Source: §4 · p. 8
  • Green hydrogen faces deployment limits because electrolysis is typically below 50% efficient and has high capital and operating costs.Most current hydrogen production instead uses carbon-intensive fossil-based routes.Source: §4 · p. 9
  • Energy-aware AI must account for both hyperscale data centers and edge systems constrained by power, memory, real-time operation, and hardware efficiency.The section calls for collaboration among mechanical engineers, energy planners, data-center designers, and policymakers on system-level solutions.Source: §4 · p. 9
  • Data centers may use a portfolio of renewables, efficiency, storage, advanced cooling, hydrogen, and other low-carbon sources tailored to regional conditions.Aluminum-based hydrogen generation is one investigated option, with no external electrical input and no direct carbon emissions reported for the process.Source: §4 · p. 9
  • Brazil’s low-carbon electricity mix and renewable potential may attract AI infrastructure, while high ambient temperatures and grid disparities complicate siting.The section makes this opportunity conditional on integrated energy planning, thermal management, and sustainability considerations.Source: §4 · p. 9

5 Physical knowledge and reasoning for better AI systems in engineering

Safety-relevant mechanical engineering tasks face data scarcity, distribution shifts, and demands for traceability and physical consistency. The section argues for combining learning with structured knowledge and physical principles, while identifying deployment challenges and research directions.

5 Physical knowledge and reasoning for better AI systems in engineering

  • Data scarcity, distribution shifts, traceability, and physical consistency motivate combining learning with domain knowledge in safety-relevant mechanical engineering.The section frames these needs in the context of high-consequence engineering tasks that cannot rely solely on data-driven pattern recognition.Source: §5 · pp. 9–10
  • Neuro-symbolic AI combines neural-network pattern recognition with symbolic inference, but integrating physical models efficiently remains difficult amid sparse, noisy data.Current models may also lack the speed and robustness needed for real-time tasks, while black-box behavior complicates safety, certification, and compliance.Source: §5 · p. 10
  • Physics-informed and neuro-symbolic methods may reduce computational cost and energy use by narrowing the hypothesis space and limiting brute-force search and retraining.The stated potential is particularly relevant when data are scarce, reliable generalization is needed, or training and inference face resource constraints.Source: §5 · p. 10
  • Proposed research directions include embedding governing equations in learning, developing mechanical-engineering ontologies, and separating neural perception from symbolic reasoning.The section also calls for real-time inference, benchmarking and standardization, and stronger causal reasoning, contextual awareness, and planning.Source: §5 · p. 10

6 Learning-based model development under physical constraints

For mechanical systems, model development and diagnostics involve inferring hidden properties from indirect measurements, often in ill-posed settings. The section presents physics-informed learning as a way to combine data-driven flexibility with physical structure, supporting more robust and interpretable models while addressing transfer and uncertainty.

6 Learning-based model development under physical constraints

  • Model development constructs mathematical representations linking system structure, parameters, and dynamics to observable behavior; diagnostics is treated as model development under partial observability and constraints.The inference is often inverse and ill-posed, lacking uniqueness or stability.Source: §6 · p. 10
  • Machine learning can approximate unknown mappings and extract patterns from heterogeneous sensor data, but purely data-driven surrogate reliability depends on representative data and relevant correlations.These surrogates rely on empirical regularization rather than physical identifiability and are effective within their domain of validity.Source: §6 · p. 11
  • Physical laws constrain admissible models, mitigating non-uniqueness and instability while supporting robustness, extrapolation, interpretability, and trust.Physics also grounds parameters and latent variables semantically, supporting validation, model reuse, and causal reasoning in diagnostics.Source: §6 · p. 11
  • Hybrid frameworks use physics-based models as structure and machine learning for residual effects, unmodeled dynamics, or uncertainty.For digital twins and long-horizon prediction, physics-informed structure supports stability, interpretability, and meaningful uncertainty quantification; multi-fidelity methods combine information across accuracy and cost levels.Source: §6 · p. 11
  • The Sim2Real gap arises when simulated training data differ from physical deployment conditions, making transferability a reliability concern.Physics-informed learning, experimental validation, uncertainty quantification, and continuous model updating are identified as mechanisms to improve transfer.Source: §6 · p. 11

7 Representative mechanical engineering applications

The representative applications span mechanical engineering from manufacturing and structural integrity to maintenance, computational analysis, materials design, and scientific workflows. Across these areas, the article emphasizes combining AI with physical models, uncertainty assessment, and engineering judgment.

7 Representative mechanical engineering applications

  • AI applications span manufacturing, structural integrity, maintenance, computational analysis, system dynamics, materials design, and scientific workflows, while other ME areas remain outside this representative survey.The selected examples cover the ME life cycle from material and component design through operation, maintenance, and discovery.Source: §7 · p. 12

7.1 Additive manufacturing

  • Melt-pool monitoring and defect detection can support process optimization, but reliable analysis must address physical mechanisms, limited data, and uncertainty.Surrogate modeling, neural operators, and hybrid simulation frameworks may accelerate exploration while preserving consistency with governing physics.Source: §7.1 · p. 12

7.2 Structural health monitoring, fatigue, and fracture

  • Structural monitoring infers partially observable damage from indirect measurements, and safety-critical deployment requires physical consistency, uncertainty-aware inference, and traceability.AI-assisted crack detection and fatigue-life prediction combine ML with fracture mechanics, while uncertainty quantification supports maintenance and inspection decisions under variable loading.Source: §7.2 · pp. 12–13

7.3 Tribology and predictive maintenance

  • Tribological applications use operational sensor data for condition monitoring and maintenance, but hybrid physical-ML methods are needed to address limited generalization.Uncertainty quantification, physical consistency, and engineering judgment help turn data into actionable decisions.Source: §7.3 · p. 13

7.4 Dynamics, control, robotics, and digital twins

  • AI supports system identification, control, anomaly detection, and digital twins, but simulation-trained models can degrade under real operating conditions.Digital twins integrate physical systems with updated computational representations; bridging the Sim2Real gap involves domain adaptation, transfer learning, hybrid modeling, and experimental validation.Source: §7.4 · p. 13

7.5 Computational mechanics and scientific computing

  • AI methods accelerate computational mechanics through surrogates, reduced-order models, operator learning, and solver acceleration while retaining numerical methods and physical principles.Neural operators and surrogates approximate solution operators for rapid parametric evaluation; uncertainty methods address model-form error and decision robustness. Equation discovery uses data to infer interpretable governing equations [27].Source: §7.5 · pp. 13–14

7.6 Materials design, multiscale modeling, and topology optimization

  • Probabilistic ML can accelerate materials exploration and quantify confidence, while data-driven microstructure design and topology optimization support constrained structural design.Physical knowledge—including constitutive behavior, conservation laws, manufacturing constraints, and multiscale interactions—remains essential to development and validation.Source: §7.6 · p. 14

7.7 Scientific discovery and autonomous engineering workflows

  • Agentic engineering workflows coordinate multi-step computational tasks, but their effectiveness depends on human-developed reasoning structures, validation, and professional oversight.These workflows may combine simulations, optimization, data analysis, and design iteration; successful computer-aided engineering workflows rely on rigorous reasoning templates and domain knowledge.Source: §7.7 · pp. 14–15

8 Knowledge formation, learning, and some limits of AI

The section distinguishes computational AI from human cognition and frames knowledge as something developed through education, observation, research, and experience. It warns that overdependence on AI-mediated searches may constrain these practices, while arguing that knowledge generation and validation remain human activities.

8 Knowledge formation, learning, and some limits of AI

  • In the paper's operational view, intelligence is processing information and applying it to solve problems, while knowledge for AI is encoded in accumulated datasets.The section asks who will provide that knowledge over the medium and long term.Source: §8 · p. 15
  • Human knowledge develops through observation, study, research, and experience, with judgment shaped by context, rational and emotional processes, and learning from errors.Motivation can also shape learning trajectories.Source: §8 · p. 15
  • Machine learning extracts statistical patterns from data to build models or make predictions, without implying human cognition or understanding.The article contrasts this computational process with lifelong human learning through education, observation, experimentation, and evaluation.Source: §8 · p. 15
  • Overreliance on AI-facilitated searches may reduce interest in formal education and limit opportunities for observation, experimentation, and critical evaluation.The section maintains that generating, interpreting, and validating knowledge remain human activities that must continue to be cultivated.Source: §8 · p. 16

9 Education and curriculum modernization

The section calls for engineering education to combine AI and data-driven methods with physical engineering fundamentals, ethics, and human-agent collaboration. It also emphasizes practical experience, ongoing professional learning, and critical verification of AI-assisted work.

9 Education and curriculum modernization

  • AI-enabled ME education should integrate engineering fundamentals, AI methods, ethical responsibility, and human-agent teaming to support critical and accountable use.The framework includes supervising workflows, inspecting AI-generated analyses, and validating computational results.Source: §9 · p. 16, Figure 4 · p. 17
  • Curricula should connect AI, mechanical engineering, and systems thinking while teaching hybrid modeling, physics-informed learning, and AI limitations such as uncertainty and bias.Hands-on projects using real datasets, experiments, and simulations help students connect AI techniques with practical engineering and domain knowledge.Source: §9 · p. 16
  • AI literacy includes recognizing automation bias, evaluating outputs critically, and maintaining accountability while preserving scientific rigor and professional responsibility.The section distinguishes this broader literacy from technical proficiency alone.Source: §9 · p. 16
  • Engineers working with agentic systems need to set objectives, supervise workflows, inspect generated code and reasoning, and check results against physical laws and requirements.Laboratory work, prototyping, instrumentation, and testing remain indispensable for developing physical intuition, practical judgment, and understanding of uncertainty.Source: §9 · p. 17
  • Continuing education through training programs, short courses, and micro-credentials is presented as essential for practicing engineers to keep pace with evolving AI technologies.A UK undergraduate survey found that the vast majority reported using generative AI for study tasks.Source: §9 · pp. 16–17

10 Ethics, responsibility, and professional accountability

The section treats ethical responsibility as inseparable from technical responsibility because AI-assisted engineering work can affect safety, reliability, economic outcomes, and public trust. It maintains that accountability remains human and calls for verification, oversight, and institutional governance.

10 Ethics, responsibility, and professional accountability

  • AI-generated outputs can appear authoritative yet be incorrect, creating ethical risk where engineering decisions require correctness rather than plausibility.The section describes hallucination as a structural consequence of statistical optimization.Source: §10 · pp. 17–18
  • Engineers remain accountable for AI-assisted decisions, designs, and analyses, including oversight of workflow objectives, assumptions, intermediate decisions, and validation.AI tools cannot be moral or legal agents, and delegating judgment can undermine due diligence and public safety.Source: §10 · p. 18
  • Over-reliance under time or cost pressure can lead to misuse of misleading or incorrect AI-generated analyses, making verification and traceability important organizational practices.Such misuse need not result from malicious intent.Source: §10 · p. 18
  • Institutional governance should include acceptable-use policies, documentation of AI-assisted workflows, validation procedures, and explicit oversight in high-consequence applications.Responsible AI also intersects with privacy, environmental impact, and equitable access.Source: §10 · p. 18
  • The article calls for Brazilian regulators, education authorities, funding agencies, and universities to coordinate guidance, workforce development, and curricular modernization.It frames ethics as a core part of professional practice and the engineer as the bearer of judgment and responsibility.Source: §10 · p. 18

11 Validation, uncertainty quantification, and certification

Trustworthy engineering AI requires more than benchmark accuracy: validation, uncertainty quantification, and certification must establish evidence suited to the model’s purpose and consequences. The evidence must address physical consistency, operating conditions, and reliability across the system life cycle.

Validation

  • Validation must test realistic operating conditions, robustness to perturbations, generalization beyond training datasets, and consistency with physical principles and engineering knowledge.Predictive performance on historical data alone is insufficient.Source: §11 · pp. 18–19

Uncertainty quantification

  • Uncertainty can arise from measurement error, incomplete data, unknown parameters, model approximations, numerical discretization, or changing operating conditions, and is classified as aleatory or epistemic.Aleatory uncertainty reflects inherent variability; epistemic uncertainty reflects incomplete knowledge, limited data, or model-form inadequacy.Source: §11 · p. 19
  • Probabilistic modeling, Bayesian inference, multi-fidelity and surrogate methods, and conformal prediction can estimate uncertainty and support risk-informed decisions.Conformal prediction offers statistically valid prediction intervals with minimal distributional assumptions.Source: §11 · p. 19

Certification

  • Certification must address training data, model updates, and operational environments through life-cycle monitoring, traceability, documentation, uncertainty reporting, and human oversight.A common safety-critical challenge is demonstrating reliable, auditable, and robust performance throughout the operational life cycle.Source: §11 · p. 19

Validation and deployment scope

  • Operational AI in safety-critical settings requires stronger evidence of robustness, reliability, traceability, and performance under uncertainty than exploratory research use.Validation should be commensurate with the consequences of using the model.Source: §11 · p. 19

12 Conclusions

The conclusions call for principled AI integration in mechanical engineering, combining physical knowledge, validation, energy awareness, and human judgment. They also emphasize professional accountability, educational renewal, and governance as AI becomes part of engineering workflows.

Principled integration

  • AI should remain a component of engineered systems, not a substitute for engineering reasoning; high-consequence outputs should be physically constrained, validated against domain knowledge, and transparently embedded in decisions.Physics-informed and hybrid learning are presented as structural requirements for reliability, interpretability, and extrapolation beyond training regimes.Source: §12 · pp. 19–20

Energy sustainability

  • AI computation’s energy and infrastructure footprint belongs within engineering judgment, making energy-aware design, efficiency, and system-level optimization part of responsible deployment.Mechanical engineers have relevant expertise in energy conversion, thermal management, and efficiency.Source: §12 · p. 20

Knowledge and human expertise

  • Responsible AI must preserve human observation, experimentation, mentorship, and critical reasoning because overreliance on automated outputs risks narrowing intellectual engagement and weakening knowledge generation.AI can amplify access to information but cannot replace discernment, contextual understanding, or ethical judgment.Source: §12 · p. 20

Accountability and education

  • Professional responsibility remains human: legitimate AI deployment requires transparent use, traceable workflows, explicit oversight, and institutional governance, alongside updated curricula and professional guidelines.The article calls for clear AI codes of conduct and treats curriculum modernization as urgent.Source: §12 · p. 20

Integrating principles

  • The proposed principles combine human-agent supervision, physical constraints, energy-aware computation, consequence-matched validation and uncertainty monitoring, human accountability, and educational renewal.Engineering education should integrate AI literacy, critical reasoning, and ethical reflection as core competencies.Source: §12 · p. 20

Recommendations

  • The recommendations span immediate governance and verification, medium-term research on hybrid and computationally efficient methods, and long-term institutional adaptation.With rigor and foresight, AI may expand analytical reach, shorten development cycles, improve decision support, and enable new scientific and technological innovation.Source: §12 · p. 20

Conclusion

  • AI should complement engineering judgment through coherent integration of human expertise, physical understanding, and computational intelligence, rather than technological acceleration alone.The conclusion identifies principled integration as the defining challenge of the coming decade.Source: §12 · pp. 19–20

Funding

The authors acknowledge support from Brazilian and U.S. funding organizations. They report no conflicts of interest or datasets generated or analyzed, and state that AI tools assisted with language polishing and image generation.

Funding

  • The work acknowledges support from CNPq, FAPERJ, CAPES, the Araucaria Foundation of Paraná, FUNDEP–ROUTE 2030, and the U.S. Office of Naval Research.The passage lists multiple grants and projects for the authors.Source: Funding · p. 21

Declarations

  • The authors declare no conflict of interest and report that no datasets were generated or analyzed and no code is applicable.Source: Declarations · p. 21
  • AI tools, including ChatGPT and Grammarly, assisted with language polishing, stylistic improvements, and iterative image generation; the authors assume responsibility for the manuscript’s intellectual content and final version.Source: Declarations · p. 21

Related papers

Loading 2610.33105v1…