Parameter estimation for land-surface models using Neural Physics

Claire E. Heaney; Maarten van Reeuwijk; Ruiyue Huang

arxiv: 2505.02979 · v3 · submitted 2025-05-05 · ⚛️ physics.ao-ph · cs.LG

Parameter estimation for land-surface models using Neural Physics

Ruiyue Huang , Claire E. Heaney , Maarten van Reeuwijk This is my paper

Pith reviewed 2026-05-22 16:26 UTC · model grok-4.3

classification ⚛️ physics.ao-ph cs.LG

keywords parameter estimationland-surface modelNeural Physicsinverse modelingsoil temperaturedifferentiable physicsflux tower data

0 comments

The pith

A Neural Physics approach estimates land-surface model parameters from soil temperature at two depths.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops an inverse modelling method for a simple land-surface model by expressing its equations in a Neural Physics framework to enable gradient-based parameter optimization. The approach minimizes the difference between model predictions and data without needing adjoint equations or any training of the model. Synthetic experiments show that single-depth temperature data cannot yield reliable parameter estimates, but data from two depths can, except that latent and sensible heat fluxes remain indistinguishable. When applied to Phoenix urban flux tower observations with fixed albedo, the method reliably recovers thermal conductivity, volumetric heat capacity, and the combined sensible-latent heat transfer coefficient. This keeps the forward model entirely physics-based while using machine-learning tools solely for differentiability.

Core claim

The governing equations of the land-surface model are embedded in a differentiable framework to allow direct optimization of its parameters by matching predictions to observed soil temperatures. Reliable estimates of thermal conductivity, volumetric heat capacity and combined heat transfer coefficient are obtained from two-depth measurements in both synthetic tests and real Phoenix data, while single-depth observations prove insufficient and latent versus sensible fluxes cannot be separated.

What carries the argument

The Neural Physics representation of the land-surface model equations that provides differentiability for gradient-based minimization of the data mismatch.

Load-bearing premise

The simple land-surface model structure is assumed to be an adequate representation of the real system so that the optimized parameters have physical meaning rather than merely compensating for model error.

What would settle it

A comparison showing that the estimated parameters deviate substantially from independently measured soil properties at the Phoenix site would falsify the physical interpretability of the estimates.

read the original abstract

We propose a novel inverse-modelling approach which estimates the parameters of a simple land-surface model (LSM) by assimilating data into a differentiable physics-based forward model. The governing equations are expressed within a machine-learning framework using the Neural Physics approach, allowing direct gradient-based optimisation of time-dependent parameters without the need to derive and maintain adjoint formulations. The model parameters are updated by minimising the mismatch between model predictions and synthetic or observational data. Although differentiability is achieved through functionality in machine-learning libraries, the forward model itself remains entirely physics-based and no part of either the forward model or the parameter estimation involves training. In order to test the approach, a synthetic dataset is generated by running the forward model with known parameter values to create a time series of soil temperature that serves as observations for an inverse problem in which the parameters are assumed unknown and subsequently estimated. We show that it is not possible to obtain a reliable estimate of the parameters using a time series of soil temperature observed at a single depth. Using measurements at two depths, reliable parameter estimates can be obtained although it is not possible to differentiate between latent and sensible heat fluxes. We also apply the approach to urban flux tower data from Phoenix, United States, and show that the thermal conductivity, volumetric heat capacity and the combined sensible-latent heat transfer coefficient can be reliably estimated whilst using an observed value for the effective surface albedo.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This shows how to make a basic land-surface model differentiable with Neural Physics for gradient-based parameter fitting from temperature data, but the real-data claims hinge on the untested idea that the simple model has no big structural errors.

read the letter

The key takeaway is that this approach lets you calibrate a physics-only land-surface model for parameters like thermal conductivity using gradient-based optimization, without deriving adjoints or training any neural networks. It recovers parameters from synthetic data at two depths and produces reasonable estimates from Phoenix flux tower observations. They do a clean job of embedding the governing equations into a machine learning framework just for the automatic differentiation part. The forward model stays fully physics-based, and the optimization minimizes the mismatch to observed soil temperatures. This avoids the maintenance issues that come with custom adjoints, especially for time-dependent problems. The synthetic experiments show that single-depth data isn't enough but two depths allow recovery of the three free parameters, though latent and sensible heat fluxes can't be separated. The main limitation is the assumption that the simple model structure captures the real system well enough for the parameters to have physical meaning. In the real-data case, any unmodeled processes could be absorbed into the fitted values, making them effective rather than truly physical. The abstract doesn't include error bars, convergence checks, or tests against model misspecification, which leaves the reliability of the Phoenix results a bit open. The stress-test point about structural error is fair here; the synthetic runs don't address it. This paper is aimed at researchers in land-surface modeling, urban climate, or data assimilation who need practical ways to tune parameters in physics models. Someone looking for alternatives to ensemble Kalman filters or adjoint methods would find it useful. I think it deserves peer review. The core method is novel in this context and the experiments illustrate the idea, though revisions could strengthen the validation side.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a novel inverse-modelling approach for estimating parameters of a simple land-surface model by embedding its governing equations in a differentiable physics-based forward model via the Neural Physics framework. This enables direct gradient-based optimization of time-dependent parameters to minimize mismatch with soil temperature observations, without deriving adjoints or training any neural network components. Synthetic experiments show that parameters cannot be reliably recovered from single-depth temperature time series but can be from two depths, although latent and sensible heat fluxes remain indistinguishable. Application to Phoenix urban flux-tower data yields estimates for thermal conductivity, volumetric heat capacity, and the combined sensible-latent heat transfer coefficient when surface albedo is fixed to an observed value.

Significance. If the central claims hold, the work demonstrates a practical route to parameter estimation in physics-based land-surface models that exploits automatic differentiation from machine-learning libraries while keeping the forward model entirely physics-based. It provides clear evidence of identifiability limits with single-depth data and produces plausible parameter values from real observations. The absence of training and the preservation of the original physics equations are notable strengths. However, the significance for physical interpretation of the Phoenix results is limited by the lack of uncertainty quantification and tests against structural model error.

major comments (2)

[Abstract and Phoenix application] Abstract and Phoenix application section: the claim that thermal conductivity, volumetric heat capacity and the combined sensible-latent heat transfer coefficient 'can be reliably estimated' from Phoenix data is not accompanied by quantitative error bars, convergence diagnostics, or sensitivity tests to model-structure error. Without these, it is unclear whether the optimized values reflect physical quantities or compensate for unrepresented processes such as vertical heterogeneity or vegetation effects.
[Synthetic experiments] Synthetic experiments section: while the perfect-model tests confirm identifiability when two depths are observed, they supply no evidence on the adequacy of the simple LSM structure for real systems. This assumption is load-bearing for the physical interpretability asserted in the Phoenix experiment, yet no cross-validation against independent fluxes or comparison to more complex models is reported.

minor comments (1)

[Method] Notation for the combined sensible-latent heat transfer coefficient should be defined explicitly at first use and used consistently in equations and text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments on our manuscript. We address each major comment below and outline the revisions we will make to strengthen the presentation and interpretation of our results.

read point-by-point responses

Referee: [Abstract and Phoenix application] Abstract and Phoenix application section: the claim that thermal conductivity, volumetric heat capacity and the combined sensible-latent heat transfer coefficient 'can be reliably estimated' from Phoenix data is not accompanied by quantitative error bars, convergence diagnostics, or sensitivity tests to model-structure error. Without these, it is unclear whether the optimized values reflect physical quantities or compensate for unrepresented processes such as vertical heterogeneity or vegetation effects.

Authors: We agree that the current presentation of the Phoenix results would benefit from additional diagnostics to support interpretability. In the revised manuscript we will add convergence diagnostics (optimization loss curves and parameter trajectories) and sensitivity tests to the fixed albedo value and vertical discretization. We will revise the abstract and Phoenix section to state that the parameters 'are estimated' from the observations under the model assumptions, rather than claiming they 'can be reliably estimated'. A new paragraph will explicitly discuss the possibility that the estimates compensate for unrepresented processes such as vertical heterogeneity or vegetation effects, thereby addressing the concern about physical interpretation. revision: partial
Referee: [Synthetic experiments] Synthetic experiments section: while the perfect-model tests confirm identifiability when two depths are observed, they supply no evidence on the adequacy of the simple LSM structure for real systems. This assumption is load-bearing for the physical interpretability asserted in the Phoenix experiment, yet no cross-validation against independent fluxes or comparison to more complex models is reported.

Authors: The synthetic experiments are intentionally constructed under perfect-model conditions to isolate the question of parameter identifiability, which is a necessary first step before real-data applications. We acknowledge that these tests do not validate structural adequacy for real systems. In revision we will add a limited cross-validation for the Phoenix case by comparing the optimized model's predicted surface sensible and latent heat fluxes against the independent flux-tower observations. A full comparison against more complex land-surface models lies outside the scope of the present work but will be noted as an important direction for future research; we will expand the discussion section to highlight the structural assumptions of the simple LSM and their implications for physical interpretability. revision: partial

Circularity Check

0 steps flagged

No significant circularity in parameter estimation via external data mismatch

full rationale

The paper's core chain is a standard inverse problem: express the LSM forward model in differentiable Neural Physics form, then optimize parameters by minimizing mismatch to independent temperature observations at one or two depths (synthetic or Phoenix flux-tower data). Synthetic recovery tests confirm identifiability under perfect-model conditions but do not redefine any output as an input. Real-data estimates fit the three parameters (thermal conductivity, volumetric heat capacity, combined heat transfer coefficient) against external measurements while holding albedo fixed; nothing reduces a claimed prediction to a quantity already fitted inside the same equations. No self-citations, uniqueness theorems, or ansatzes appear as load-bearing steps. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that the chosen simple LSM equations are a sufficient forward operator and that the observed temperatures provide enough independent information to constrain the target parameters. No new physical entities are introduced.

free parameters (3)

thermal conductivity
Target parameter being estimated; its value is adjusted by the optimizer to minimize temperature mismatch.
volumetric heat capacity
Target parameter being estimated from the data.
combined sensible-latent heat transfer coefficient
Target parameter recovered from Phoenix observations.

axioms (2)

domain assumption The governing equations of the simple land-surface model accurately represent the dominant heat-transfer processes at the measurement sites.
Invoked when the authors interpret the optimized values as physically meaningful rather than effective parameters that absorb model error.
domain assumption Temperature observations at the chosen depths or flux-tower locations are sufficient to constrain the parameters without requiring additional regularization or prior distributions.
Required for the claim that reliable estimates are obtained from two-depth or Phoenix data.

pith-pipeline@v0.9.0 · 5781 in / 1600 out tokens · 76016 ms · 2026-05-22T16:26:09.908725+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

J(p) = 1/|I||N| Σ (T(xi,tn;p) − T̃(xi,tn))² ; p* = arg min J(p) using Adam
IndisputableMonolith/Foundation/DimensionForcing.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

C ∂T/∂t = λ ∂²T/∂x² with surface energy balance and two-depth temperature assimilation

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.