Freezing the Weather Pathway: Parameter Isolation for Climate-Penalty Attribution

The "climate penalty" is the extra air pollution that a warmer, sunnier, more stagnant atmosphere produces even when emissions stay flat or fall. It matters because it can cancel clean-air policy. In the Yangtze River Delta in 2022, meteorology pushed ozone up by 9.2 μg/m³ while emission controls cut it by only 1.5 μg/m³ (Zhao et al.). Measuring the penalty currently takes paired runs of chemical transport models, the physics-based simulators of atmospheric chemistry and wind transport. One run uses real emissions and a second holds emissions fixed. These runs take hours per city, so studies fall back on short time slices (Cui et al.). Fast neural forecasters exist, but they mix weather and emission information in shared weights. They do not say how much of a trend came from the weather.

We propose to build a neural air-quality model with two separate parameter blocks, one for the weather response and one for the emission-driven trend, and to train them in a fixed order. The weather block is fit first on short-timescale variation. Within a month, emissions barely change, so day-to-day swings in ozone are mostly meteorological. That block is then frozen, so its weights cannot change. The emission block is fit afterwards on the multi-year series. The freeze is the key ingredient. Over a decade, warming and nitrogen-oxide cuts trend together. A model trained jointly on that data is free to credit emission-driven changes to temperature or the reverse. Freezing the weather block after it has been fit on data where emissions are nearly constant stops the confounded trend data from rewriting it. The climate penalty can then be read off the frozen block by feeding it new and old meteorology.

Core Claim

Continual-learning methods use parameter isolation to stop a model from forgetting old tasks while it learns new ones. We claim the same machinery serves a different purpose here. Parameter isolation is a tool for keeping causal pathways identifiable. If the weather-response parameters are learned only from data where emissions are roughly constant, and are then shielded from later updates, the model's counterfactual "same emissions, different weather" answers should match a chemical transport model's fixed-emission runs much more closely than answers from a model of the same capacity trained jointly. We test this claim in a setting where the true penalty is known.

Relation to Prior Work

The atmospheric literature motivates the problem and supplies the ground truth. Factorial chemical transport model studies show that the penalty is large and nonlinear. Lu et al. attribute 25–31% of the 1995–2019 surface ozone rise in East and Southeast Asia to climate. Their ML components, which are random forests and ridge regression, are static fits with no structural separation of the two pathways. Zhao et al. find that ozone becomes hypersensitive once low cloud cover drops below 0.3.

Two papers constrain the design. Single-variable perturbation runs of the chemical transport model CMAQ give regional temperature sensitivities, such as 1.1 ppb/K for July ozone in Beijing (Shi et al.). They also show that the sign of the heat effect flips between polluted and clean chemical regimes (Shi et al.). The weather block therefore cannot be emission-blind. It must be conditioned on a slowly varying descriptor of the chemical regime.

Neural forecasters set the speed bar but leave attribution open. FuXi-Air fuses AI weather forecasts, emission inventories and station data through cross-attention and shared weights. It reports 36–67% lower RMSE than the WRF-CMAQ chemical transport model for O₃, SO₂ and CO in Shanghai, and its ablations show meteorology dominates accuracy. That dominance is exactly why its weights cannot be trusted to keep the two pathways apart. AirPhyNet builds advection and diffusion into a graph network but explicitly leaves out photochemistry. The photochemistry is where the penalty lives. Air Quality Arena shows that zero-shot foundation models forecast well, but it is univariate and cannot represent a weather counterfactual at all.

The continual-learning literature supplies the method. Winning SubNetworks freeze allocated weights and guarantee zero backward transfer, meaning later training cannot degrade earlier tasks. Functional Task Networks show the guarantee only holds if the optimizer's momentum states are reset between stages. Without the reset, frozen-looking weights still drift. SPG scales gradients down rather than zeroing them, and we use it as a soft-isolation variant. Additive parameter decomposition splits weights into shared and task-specific parts, which is the template for our block structure. None of these papers treats the isolated blocks as causal pathways or evaluates them on an attribution target. That is the departure we make.

Research Program

The model predicts hourly or daily-maximum 8-hour ozone at station $s$ and time $t$ as

$$\hat{y}_{s,t} = f(m_{s,t}, r_{s,t};\, \theta_M) + g(e_{s,t}, s, t;\, \theta_E),$$

where the inputs and parameter blocks are as follows:

  • $m$ is the vector of local and upstream meteorology from ERA5: temperature, wind, humidity, boundary-layer height, low cloud cover and surface shortwave radiation.
  • $r$ is a slowly varying chemical-regime descriptor, built from inventory ratios of nitrogen oxides to volatile organic compounds plus monthly-mean NO₂.
  • $e$ is the gridded emissions from MEIC or CAMS.
  • $\theta_M$ is the weather block and $\theta_E$ is the emission/trend block. The two share no parameters.

The weather block is a small graph network over stations, so that upwind meteorology can matter.

Training runs in three stages.

Stage 1: fit the weather block. We fit $\theta_M$ on within-month anomalies of ozone against within-month anomalies of meteorology. A per-station, per-month offset absorbs everything slow.

Stage 2: freeze and fit the emission block. $\theta_M$ is frozen with a hard mask, its gradients are zeroed, and the optimizer state is reset. $\theta_E$ is then fit to the full series.

Stage 3: optional soft refinement. Both blocks are refined jointly, but SPG-style gradient scaling limits how far $\theta_M$ can move.

The climate penalty between two periods is estimated as $f(m_{\text{new}}, r) - f(m_{\text{old}}, r)$ with $r$ held at a reference value. This is the neural analogue of a fixed-emission factorial run, computed in seconds.

The program has a secondary continual-learning use that the same structure supports. As new years arrive, only $\theta_E$ is updated. If a meteorological regime appears that is absent from history, such as low cloud cover persistently below 0.3, a new isolated subnetwork is added to $\theta_M$ rather than overwriting the old one. We treat this as an extension, not the core claim.

The obvious weakness is transfer across timescales. Weather sensitivity learned from day-to-day swings may not equal sensitivity to a decade-long climatological shift. For example, synoptic stagnation differs from a long-term decline in cloud cover. The experiments below test this directly rather than assuming it away.

Baselines

The naive reference is a station-wise seasonal climatology plus linear trend. For attribution, it is multiple linear regression of ozone on meteorology and emissions, with the penalty read from the meteorology coefficients, as in the statistical arm of Lu et al.. Any neural method must beat this to justify itself.

The standard ML attribution practice is a random forest weather-normalization baseline. The forest is trained on all inputs, and the penalty is computed by swapping in alternative meteorology at prediction time. This mirrors the random-forest component of Lu et al. No reported attribution error for that model against counterfactual chemical transport runs was found on Lacuna. It is the method a domain reviewer will expect us to beat.

The strongest competitor is a monolithic multimodal network in the style of FuXi-Air. It uses cross-attention over the same meteorology, emission and station inputs, is matched to our model in parameter count, and computes the penalty by the same input swap. FuXi-Air's reported result is a 36–67% RMSE reduction against WRF-CMAQ for O₃, SO₂ and CO in Shanghai (Geng et al.). It is a forecasting number, not an attribution number. We do not have FuXi-2.0 forecasts, so we re-implement the architecture on ERA5 inputs rather than claim to reproduce that figure.

The ablation that isolates our key ingredient is the identical two-block architecture trained jointly from scratch, with no anomaly stage and no freeze. A second ablation keeps the anomaly-then-trend schedule but replaces the hard freeze with ordinary fine-tuning. A third replaces it with elastic weight consolidation, which penalizes changes to weights estimated as important for the earlier stage.

Fairness. All models see the same stations, years and inputs. Each gets the same hyperparameter search budget, which is 40 configurations tuned on a held-out validation year for forecast loss only. The attribution target is never used for tuning. All models also share the same GPU budget.

What counts as success. Beating a baseline means at least 30% lower error against the true fixed-emission penalty in the twin experiment described below, with forecasting RMSE no more than 5% worse than the monolithic network.

Why not just…?

Why not just use a single strong network and swap the meteorology at prediction time? This is the default, and it fails exactly when it matters. When warming and emission cuts co-trend, nothing in the joint loss forces the network to put the trend in the right input. An input swap can return a confident but wrong penalty. FuXi-Air's own ablation, in which meteorology dominates accuracy (Geng et al.), is consistent with this risk. The twin experiment measures it directly. If the monolithic network's swapped penalty matches the true counterfactual as well as ours does, the proposal fails, and we will report it that way.

Why not just build the additive two-block architecture and skip the staged freezing? This is the most serious threat. Structural separation alone might be enough, and parameter isolation would then be decoration. The joint-training ablation exists to settle this question. Our prediction is that joint training lets $\theta_E$ absorb part of the warming trend through its time input, or lets $\theta_M$ absorb part of the emission trend. We prediction it does so to a degree that grows with the correlation between the two trends. We will vary that correlation in the synthetic setting to test the prediction.

Why not just fine-tune with a small learning rate or a weight penalty such as elastic weight consolidation? Under concept shift, soft constraints leak. In the LLM style-transfer stream, LoRA fine-tuning forgot +7.37 nats while isolation forgot +0.07 (Functional Task Networks). Leakage here does not cost accuracy on an old task. It moves the pathway that defines the attribution. We still run both variants, along with SPG soft masking, because a mild leak might help when the timescale-transfer assumption is imperfect.

Why not just train an emulator on existing fixed-emission chemical transport runs? Where such runs exist, this is a real alternative. It inherits the simulator's biases, though, and those biases are large: the three models in Lu et al. span a 6-point range in climate share. It also does not scale to regions without the runs. We will report how our observation-trained estimate falls relative to that multi-model spread. A multi-fidelity variant, which pretrains $\theta_M$ on simulator runs and then freezes it, is a natural follow-up that the same isolation machinery supports.

Proposed Experiments

Twin experiment (decisive test). We use a decade of WRF-Chem or CMAQ output over the Yangtze River Delta and the North China Plain, with time-varying MEIC emissions, plus the matching fixed-emission counterfactual run. These follow the configurations of Zhao et al. and Lu et al.; obtaining the runs or rerunning them is the main compute cost. All models are trained only on the time-varying run, sampled at real station locations, with noise added. They are scored on how well they recover the counterfactual. The metrics are:

  • RMSE of the meteorology-induced ozone change for each station and season.
  • Bias in the regional climate share.
  • Correlation of the year-to-year penalty with the truth.

A synthetic sweep then rescales the emission trend so that it is more or less correlated with the temperature trend. This sweep tests the specific prediction that the advantage of freezing grows with that correlation.

Sensitivity check against perturbation physics. We take finite-difference temperature, wind and boundary-layer sensitivities from $\theta_M$ and compare them with the CMAQ perturbation coefficients for January and July 2013 (Shi et al.). The comparison covers magnitude, for example 1.1 ppb/K for July ozone in Beijing, and the regime-dependent sign flip (Shi et al.). Getting the sign wrong in clean regimes would falsify the regime-conditioning design.

Real-observation checks. We use national monitoring stations with ERA5 and MEIC inputs, and the Beijing/Shenzhen data released with AirPhyNet. The model must reproduce the sign and ordering of the 2022 Yangtze River Delta decomposition: a meteorological push of about 9.2 μg/m³ against an emission cut of about 1.5 μg/m³ (Zhao et al.). It must also land inside the 25–31% surface climate share for East and Southeast Asia (Lu et al.).

Forecasting skill. We report 24–72 hour forecast RMSE and MAE on the same splits to show that isolation costs little predictive skill. On the 24-hour task we also include the Seasonal Naive and PatchTST baselines from Air Quality Arena, run with meteorology added.

Timescale-transfer test. A negative result on this test would be informative. We train $\theta_M$ on anomalies from 2014–2018 and ask whether it predicts the climatological shift of 2019–2024 in the twin world. If it does not, we have measured how far synoptic sensitivity is from climate sensitivity. That gap matters for every weather-normalization study, not just ours.