{"id":"f833dd25-db90-4c5d-9d77-629f1f02feb6","arxiv_id":"2506.13777","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"This survey groups physics-informed AI methods for urban systems into three integration paradigms and seven method types, then maps them across eight urban domains.","lead":"This survey reviews how artificial intelligence and physics-based models are combined to understand cities, organizing the methods into three integration paradigms and seven technique types. It maps these methods onto eight urban domains, which makes it a useful entry point for researchers choosing a modeling approach for urban problems.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assignment of papers to the seven method types is not reproducible: Table S4 mislabels waste/battery/soil entries and reference [197] carries four different Fusion Method labels across tables, so the taxonomy's guidance is built on unreliable examples.","rationale":"The reader's weakest assumption is essentially correct: a survey whose contribution is a classifying framework needs reproducible assignments. My read of Tables S3–S9 confirms this assumption is violated by concrete, checkable inconsistencies. The most load-bearing is the repeated citation [197] with incompatible method labels; one paper cannot be Methods 5, 6, 7, and 4 simultaneously without an explanation. The misaligned Table S4 rows reinforce that the supplementary data were not carefully checked. The paper has real strengths: the taxonomy itself is clearly explained in Figures 2–5, the method descriptions in Section 2 are broadly consistent with the literature, and the application-by-application narrative is informative. But the empirical grounding for 'guides selection' is the set of representative assignments, and that grounding is currently unreliable. This does not require rejection; it requires a corrigendum and a verification pass, exactly the CONDITIONAL position. I therefore leave the reader's verdict unchanged.","tokens_in":33608,"tokens_out":4650,"duration_ms":56769,"concrete_test":"Build a citation-to-method incidence matrix from Tables S3–S9 and flag every reference that appears in multiple rows with different Fusion Method labels or different physical-model descriptions. Start with [197], [116], [119], [165], and [142]; read each source paper and record the actual integration mechanism and target domain. If [197] receives more than one method label, or if Table S4's Soil/Waste/Carbon rows still map [119], [116], and [165] to the wrong physical models, the paper-to-taxonomy assignments are not reproducible and the tables must be corrected before the claimed guidance can be accepted. If every flagged reference has a single consistent method label, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The survey's central claim is that its taxonomy clarifies degrees and directions of physics-AI integration and guides method selection. That claim stands or falls on whether the paper-to-method assignments in Tables S3–S9 are accurate and reproducible. They are not internally consistent. In Table S4, the Soil block lists [119] as 'Solid Waste System Dynamics' and [116] as 'Battery Degradation Models', while the main text (Sections 4.2.2 Soil/Waste) and the actual cited papers identify [55] and [119] with solid waste, [165] with battery degradation, and [116]/[142] with soil hydrology and consolidation; the table's rows are misaligned. More damaging, reference [197] (Zhang et al., KDD 2022, a crowd-simulation paper) is assigned Fusion Method 5 in Table S4 Water, Method 6 in Table S4 Soil, Method 7 in Table S8 Safety, and Method 4 in Table S9 Mobility, and is also cited in Section 2.7 for a subgrid closure application. A taxonomy of methods should assign one method per paper; four different labels for one paper mean the classification is not applied consistently and cannot be independently reproduced. Because Section 5.1's selection guidance is illustrated through these assigned examples, the unreliable mappings directly weaken the paper's advertised practical contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript surveys physics-informed AI (PIAI) methods for urban systems. It proposes a taxonomy of three paradigms—Physics-Integrated AI, Physics-AI Hybrid Ensemble, and AI-Integrated Physics—subdivided into seven methods (PINN with loss function, PINN with weight initialization, PINN with architecture design, sequential and parallel physics-AI ensembles, AI-discovered physics models, and neural physics models), and reviews applications across seven or eight urban domains. The paper claims that this taxonomy clarifies the varying degrees and directions of physics-AI integration and guides method selection for urban modeling tasks.","tokens_in":33897,"tokens_out":4530,"duration_ms":54584,"significance":"A reliable survey of this kind would be valuable: the literature is scattered across energy, environment, transportation, economy, information, public services, and emergency management communities, and Table S2's compilation of typical physical laws per urban domain is a useful reference service. The paper also articulates relevant open problems, including theoretical foundations, data efficiency, cross-domain transfer, and digital twin deployment. However, the contribution is primarily classificatory, and the value of the taxonomy depends entirely on the accuracy and consistency of the paper-to-method and paper-to-physical-theory assignments in Tables S3–S9. The current version contains systematic misassignments, so the central claim is not yet supported at the level of reproducibility that a survey of this kind requires.","major_comments":[{"comment":"The mapping in Table S4 between cited papers, physical models, and fusion methods is systematically misaligned with both the main text and the actual cited papers. For example, [119] is Ouyang et al. 2024 on physics-informed neural networks for laterally loaded piles, but the Soil block lists it as 'Solid Waste System Dynamics'; [116] is Oikawa and Saito 2024 on inverse analysis of soil hydraulic parameters, but the table lists it as 'Battery Degradation Models'; [165] is Tao et al. 2025 on battery degradation, but the table lists it as 'Vegetation-Soil Combustion Physics'; [55] is He et al. 2023 on solid waste management, but the Waste block lists it as 'Radar Backscatter Dynamics'; [154] is Singh and Gaurav 2024 on soil moisture estimation, but the table lists it as 'Pile-Soil Dynamics Models'; [142] is Seydi et al. 2024 on vegetation and soil burn severity, but the table lists it as 'Richards Soil Hydrodynamics'; and [180] is Xie et al. 2024 on soil temperature simulation, but the table lists it as 'Terzaghi Consolidation PDE'. These are not cosmetic typos: Table S4 is the evidence base for the environmental classification, and a reader attempting to use the taxonomy for method selection would be directly misled.","section":"Table S4 and Section 4.2.2"},{"comment":"Reference [197] (Zhang et al., KDD 2022, a crowd-simulation paper) is assigned four different fusion methods across the supplementary tables: Method 5 in Table S4 under Water ('Hydro-Mechanical PDEs'), Method 6 in Table S4 under Soil ('2D Soil Consolidation PDE'), Method 7 in Table S8 under Safety and Social Security ('Social force model'), and Method 4 in Table S9 under Mobility Simulation ('Social Force'). It is also cited in Section 2.7 as an example of replacing subgrid closure models in large-eddy simulation. A single paper cannot instantiate four method types and several unrelated physical models; the classification is therefore not applied consistently and cannot be independently reproduced. This directly weakens the Section 5.1 claim that the taxonomy, illustrated through reviewed examples, guides method selection.","section":"Tables S4, S8, S9 and Section 2.7"},{"comment":"The number of urban domains is internally inconsistent. The abstract states that the paper examines applications across 'eight key urban domains: energy, environment, economy, transportation, information, public services, emergency management, and the urban system as a whole', while Section 1 and Section 3 explicitly define seven core subsystems and Fig. 6 shows seven subsystems. Section 4.8 then treats 'the urban system as a whole' as an additional category. The abstract, introduction, and conclusion need to agree on whether the taxonomy covers seven subsystems plus one holistic category or eight co-equal domains, since the survey's coverage claim is part of its advertised contribution.","section":"Abstract, Section 1, Section 3, Section 4.8"},{"comment":"The 'AI-Discovered Physics Model' label (Method 6) is applied to papers that do not appear to discover governing equations from data. For instance, [35] (SLANT) uses labeled jump-diffusion stochastic differential equations in a probabilistic framework, [117] (SINN) transforms existing opinion dynamics models into ODEs approximated by neural networks, and [170] (ODID) adopts a heat transfer equation to model information diffusion. Labeling these as Method 6 conflates 'using a known equation in a neural network' with 'discovering a governing law', which contradicts the definition of Method 6 in Section 2.6. If Method 6 is intended more broadly, the definition should say so; otherwise the assignments in Table S7 are inconsistent with the taxonomy.","section":"Table S7 and Section 4.5"}],"minor_comments":[{"comment":"There are several typographical errors, including 'Reaserch' in the Section 4.5.1 heading and 'for for' in Section 4.7.2; a careful proofreading pass is needed.","section":"Section 4.5.1 and Section 4.7.2"},{"comment":"The spatiotemporal decay model equation contains a stray double comma and an unbalanced parenthesis in the displayed formula; the notation should be cleaned up.","section":"Section 4.7.2"},{"comment":"The method labels in the rightmost columns are written inconsistently, sometimes as 'Method 1 & 6' and sometimes as 'Method 1 & 6.' with a period, and the tables do not state whether multiple labels mean that one paper uses multiple methods or that the assignment is uncertain; this should be clarified for reproducibility.","section":"Tables S3–S9"},{"comment":"The statement that 'data-driven strategies often favor AI-dominant approaches' is a reasonable high-level observation, but it is not derived from the surveyed tables; tying this claim to a quantitative breakdown of the papers in Tables S3–S9 would strengthen it.","section":"Section 5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be a broadly scoped early survey with useful material, but the supplementary tables—the main evidence for the taxonomy—need to be rebuilt and verified against the cited papers before the manuscript can be considered for publication. I would advise the editor to require the authors to re-check every entry in Tables S3–S9, fix the inconsistent use of reference [197], and harmonize the domain count across the abstract, introduction, and conclusion. The scope is appropriate for a survey journal, but the current reliability issues are load-bearing rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you're thinking about physics-AI fusion in cities, but don't cite the tables without checking them. The three-paradigm, seven-method taxonomy is clean and does organize the space well; it's a genuinely useful framing for a field that usually reviews methods in isolation. The domain walk-through (energy, environment, transportation, economy, information, public services, emergency management) is broad and mostly sensible, and it gives a newcomer a good map of where PINNs and hybrid models have been applied.\n\nThe soft spots are real. The stress-test note is right: the supporting tables, especially Table S4, have systematic misassignments. Soil rows include solid-waste and battery-degradation entries, which the main text itself assigns elsewhere. Reference [197] (a crowd-simulation paper) is labeled as four different fusion methods in four different tables. That's not a typo; it means the method classification is not applied consistently enough to be independently reproduced. The domain count also wobbles between seven and eight, depending on whether you include 'urban system as a whole.' Section 5.1's selection advice leans on these examples, so the practical guidance inherits the table errors.\n\nThe selection of representative papers leans on the authors' own group more than a neutral reading would expect, but that's a lesser concern; the taxonomy itself is a reasonable reorganization of known methods, not a novel result.\n\nBottom line: as a survey it's a decent entry point, but as a reference work it currently needs a careful corrigendum. The core framing is fine, the execution is sloppy. I'd send it to review but require that the tables be rebuilt and the domain count reconciled before publication. Would I cite it? Only after those fixes; right now it's a map with mislabeled streets.","headline":"A genuinely useful taxonomy for physics-AI fusion in urban systems, but the supporting classification tables are so misaligned that every assignment needs verification before the framework can be trusted.","tokens_in":34410,"tokens_out":2497,"would_cite":false,"duration_ms":31416,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A survey maps physics-AI fusion for cities into seven method types.","keywords":["physics-informed AI","urban systems","taxonomy","PINN","neural ODE","SINDy","social force model","smart cities"],"falsifier":"Re-examine citation [197] across the tables: it appears as \"Hydro-Mechanical PDEs\" under water (Table S4), as \"2D Soil Consolidation PDE\" under soil (Table S4), and as \"Social force model\" under public services (Table S8). One citation cannot instantiate three different physical models, so a reproducible assignment of papers to physical theories is missing; checking such duplicates on a sample of the cited papers would determine whether the taxonomy's classifications are stable.","tokens_in":33376,"feed_emoji":"🏙️","tokens_out":4613,"duration_ms":49111,"temperature":0.7,"pith_summary":"This survey organizes the growing field of physics-informed AI for urban systems into three integration paradigms—Physics-Integrated AI, Physics-AI Hybrid Ensemble, and AI-Integrated Physics—and seven concrete method types. It claims this taxonomy captures the degree and direction of physics-AI fusion and can guide practitioners in choosing a method based on how much physical knowledge is available, how much data exists, and what spatiotemporal scale matters. It applies the taxonomy across eight urban domains from energy and transport to public health and emergency response. The payoff of a right taxonomy would be a principled way to select, compare, and combine methods for urban modeling rather than relying on ad-hoc choices.","feed_headline":"Physics and AI for cities, sorted into seven fusion types","feed_subtitle":"New taxonomy groups methods by how deeply physics and AI are integrated across energy, transport, health, and more.","key_machinery":"The organizing device is a two-level taxonomy: three paradigms based on the direction and degree of physics-AI integration, further split into seven representative method types—PINN with loss function, PINN with weight initialization, PINN with architecture design, sequential physics-AI ensemble, parallel physics-AI ensemble, AI-discovered physics model, and neural physics model. The taxonomy does not itself solve any urban problem; it works as a classification and selection instrument, and it carries the survey's claim that method choice should be driven by physical knowledge, data availability, and spatiotemporal scale.","core_discovery":"On the paper's own terms, the central contribution is a structured map of how physics and AI can be combined in urban modeling. The three paradigms are ordered by which side dominates: physics laws embedded into neural networks (Physics-Integrated AI, including PINNs with physics loss, physics-based weight initialization, and physics-guided architecture), balanced pipelines where physics and AI alternate or run in parallel and their outputs are fused (Physics-AI Hybrid Ensemble), and physics-dominated structures with AI replacing specific modules or discovering the governing equations (AI-Integrated Physics, including AI-discovered physics and neural physics models). The paper argues this classification clarifies the degrees and directions of integration, making it easier to match methods to urban applications and data conditions.","pith_inferences":["If the taxonomy is right, it implies a cost-benefit ordering among the seven methods that the paper does not make explicit: methods 1-3 trade training cost for physical consistency, while methods 4-5 trade architectural complexity for flexibility.","A testable extension would be to apply the taxonomy to a fresh batch of recently published urban PIAI papers and measure inter-annotator agreement; the paper's own Table S4 suggests such agreement may be low when the same citation is mapped to different physical models.","The taxonomy could be operationalized as a decision rule linking data quantity, physics explicitness, and chosen method, turning the survey's qualitative guidance into a falsifiable design recipe.","The paper's grouping suggests that the eight-domain matrix could be remapped onto subsystem-level physical processes, which would sharpen selection guidance for practitioners."],"forward_implications":["A practitioner can use the taxonomy to pick a starting point: physics-dominant methods when physical laws are explicit and data is scarce, AI-dominant methods when data is rich and dynamics are complex, and hybrid ensembles in between.","The taxonomy exposes gaps: the paper notes PINNs lack convergence and generalization guarantees, most applications are domain-specific, and no unified framework yet integrates heterogeneous physical laws into AI architectures.","The paper points to foundation models and world models that embed physical priors as the likely next step for urban digital twins.","Cross-domain urban modeling, such as coupling flood dynamics with traffic flow, emerges as a growth area enabled by physics-informed graph networks and neural ODEs.","The survey concludes that urban computing is a natural arena for this fusion because cities combine strong physical priors with abundant, heterogeneous observational data."],"supporting_citations":[{"why":"Supplies the original PINN framework that Method 1 (PINN with loss function) builds on.","marker":"[129]"},{"why":"Introduces SINDy, the sparse-identification basis for Method 6 (AI-discovered physics).","marker":"[137]"},{"why":"Introduces Neural ODEs, the backbone of Method 7 (neural physics models) and physics-guided architecture design.","marker":"[25]"},{"why":"Defines the social force model, a core physical law used across public services and emergency management examples.","marker":"[57]"},{"why":"Provides the general physics-informed machine learning framing that the survey's three paradigms extend.","marker":"[71]"},{"why":"Anchors Method 4 (sequential physics-AI ensemble) with a machine-learning-accelerated CFD example.","marker":"[75]"},{"why":"Anchors Method 5 (parallel physics-AI ensemble) as a hybrid strategy for systems with limited data or incomplete physics.","marker":"[189]"}],"fun_headline_variants":["City modeling gets a physics-AI fusion taxonomy","Three paradigms for merging physics and AI in cities","Seven urban methods scored by physics-AI integration","Physics-AI fusion: a map for urban system models","New taxonomy sorts physics-AI combos for cities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The taxonomy's practical guidance works only if each reviewed paper has been classified into the correct method type and paired with the correct physical theory; the survey's own tables show that some citations are assigned to different physical models in different places.","fun_headline_variants_meta":{"raw":{"variants":["City modeling gets a physics-AI fusion taxonomy","Three paradigms for merging physics and AI in cities","Seven urban methods scored by physics-AI integration","Physics-AI fusion: a map for urban system models","New taxonomy sorts physics-AI combos for cities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1434,"prompt_tokens":891,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":469}},"tokens_in":507,"tokens_out":543,"duration_ms":7166,"temperature":1.0,"reasoning_tokens":469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:32:48.550350+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-examine citation [197] across the tables: it appears as \"Hydro-Mechanical PDEs\" under water (Table S4), as \"2D Soil Consolidation PDE\" under soil (Table S4), and as \"Social force model\" under public services (Table S8). One citation cannot instantiate three different physical models, so a reproducible assignment of papers to physical theories is missing; checking such duplicates on a sample of the cited papers would determine whether the taxonomy's classifications are stable.","supporting_citations":[],"review_version":1}