LoRA Adaptation Strength in Aurora-WRF: Trade-offs in Regional Heavy-Rainfall Forecasting

Boyan Liu; Xiaoyuan Zhang

physics.ao-phSep 24, 2026

LoRA Adaptation Strength in Aurora-WRF: Trade-offs in Regional Heavy-Rainfall Forecasting

Authors: Boyan Liu, Xiaoyuan Zhang

Organizations: Zhongguancun Academy, Beijing, China

Abstract

We investigate how parameter-efficient adaptation of an atmospheric foundation model affects downstream regional precipitation in a controlled Aurora-WRF coupling experiment over the Beijing-Tianjin-Hebei region. A single-step, precipitation-weighted LoRA adaptation of AuroraPretrained is trained on May-September 2020-2021 and selected using 2022 validation data. Five inference-time adaptation strengths are coupled to an unchanged 9-km WRF configuration and compared with a GFS-driven control under common ERA5 initialization and auxiliary-field rules. Evaluation uses 17 initializations in 2023, hourly ERA5 precipitation, and conservative remapping to a fixed 356-cell, 0.25-degree verification grid. Heavy rainfall is defined exclusively as at least 50 mm in a continuous 24-hour window, advanced hourly. For window starts from 0 to 48 hours, unadapted Aurora attains the highest mean critical success index (CSI), 0.216, compared with 0.144 for the GFS control. An intermediate LoRA strength of 0.75 reduces the Aurora-driven 24-hour precipitation mean squared error by 30.7% and brings pooled frequency Bias from 1.510 to 1.024, but lowers CSI to 0.182. GFS retains the lowest whole-domain error. These results characterize a trade-off between rainfall detection, frequency calibration, and intensity error, showing the importance of evaluating adaptation strength against multiple downstream criteria within a common coupling framework.

Figures & tables

Field	Source and treatment
Soil temperature/moisture, skin temperature, SST, sea ice, snow	ERA5 at the start of the adjustment period; prescribed auxiliary fields persist. The land-surface model evolves its internal state; SST updating is disabled.
Surface pressure $p_{s}$	After $t_{0}$ , $\widehat{p}_{s}(t)=[p_{s}(t_{0})/p_{\rm msl}(t_{0})]\widehat{p}_{\rm msl}(t)$ .
2-m specific humidity $q_{2}$	Derived from ERA5 dew point and pressure at $t_{0}$ ; scaled subsequently by forecast lowest-level specific humidity relative to its initial value and capped at saturation.
Cloud liquid and ice	ERA5 cloud fields only at initialization; subsequent prescribed cloud boundary mixing ratios are zero.
Static fields	Common land–sea mask, orography, soil categories, and WPS geography.

Table 1: Shared auxiliary-field rules. Initialization fields are not future observations.

Driver	CSI	Bias	POD	FAR	MAE (mm)	MSE (mm 2 )
GFS	0.144	0.896	0.218	0.757	5.17	167.88
$\alpha=0$	0.216	1.510	0.405	0.732	7.16	302.20
$\alpha=0.25$	0.202	1.326	0.332	0.749	6.81	296.94
$\alpha=0.50$	0.192	1.131	0.305	0.730	6.40	233.14
$\alpha=0.75$	0.182	1.024	0.286	0.720	6.47	209.57
$\alpha=1$	0.180	1.649	0.327	0.802	7.01	274.73

Table 2: Main 17-initialization results, 24-hour window starts 0–48 h. Each CSI averages 49 pooled window scores. MAE and MSE use all verification cells.

Figure 1: Pooled main-scope performance for all six drivers. Bars deliberately show the established aggregate estimators without attaching case-level standard deviations to a different estimator. Case dispersion is reported separately in Table 3 and Figure 3 .

Figure 2: Skill and errors versus forecast lead. Solid curves retain the pooled CSI/Bias and mean MSE estimators. Shading and its dotted boundaries show

\pm 1

bootstrap SD for CSI/Bias and

\pm 1

case SD for MSE; these are not confidence bands. Negative lower bounds reflect the symmetric SD construction, not negative scores. For the first three panels, lead is the 24-hour window start : 0 denotes 0–24 h and 48 denotes 48–72 h. Hourly MSE uses interval ends 1–72 h, with no fabricated hour-zero error.

Driver	Case CSI	Case Bias	MSE 24 (mm 2 )	MSE 1 (mm 2 )
GFS	$0.037\pm 0.056$ (10)	$1.173\pm 1.343$ (9)	$167.88\pm 262.22$ (17)	$1.45\pm 2.01$ (17)
$\alpha=0$	$0.037\pm 0.082$ (12)	$2.523\pm 4.393$ (9)	$302.20\pm 541.73$ (17)	$2.30\pm 3.63$ (17)
$\alpha=0.25$	$0.025\pm 0.075$ (12)	$2.486\pm 5.496$ (9)	$296.94\pm 626.85$ (17)	$1.97\pm 4.02$ (17)
$\alpha=0.50$	$0.026\pm 0.077$ (10)	$5.786\pm 15.880$ (9)	$233.14\pm 376.13$ (17)	$1.53\pm 2.92$ (17)
$\alpha=0.75$	$0.025\pm 0.076$ (10)	$5.780\pm 15.427$ (9)	$209.57\pm 273.28$ (17)	$1.56\pm 2.43$ (17)
$\alpha=1$	$0.025\pm 0.079$ (11)	$19.949\pm 56.790$ (9)	$274.73\pm 390.98$ (17)	$2.00\pm 3.00$ (17)

Table 3: Initialization-level mean

\pm

sample SD; parentheses give valid initialization counts. CSI and Bias here are case statistics , not the headline pooled estimators. MSE 24 averages the 49 windows; MSE 1 averages hours 1–72.

Figure 3: Initialization-level distributions over the 17 starts. Boxes show quartiles and medians; whiskers extend to observations within 1.5 interquartile ranges. CSI and full-range Bias use symmetric-log axes (base 10, linear scale 1), linear from zero to 0.01 and 1, respectively, and logarithmic above these thresholds. Zeros and all extremes are retained; box statistics use untransformed values. Both Bias panels show the same nine defined cases per driver. The 0–6 detail remains linear; triangles mark values above its range. The dotted line is Bias=1. Counts denote valid cases, not independent events.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Date	Event cell-windows	Max event cells/window	Max 24 h (mm)
2023-05-07	0	0	0.23
2023-05-15	6	1	51.00
2023-05-21	0	0	2.01
2023-06-03	0	0	1.12
2023-06-14	0	0	1.36
2023-06-26	0	0	47.92

Appendix

Table 4: The 17 evaluation initializations (00 UTC, 2023). Reference statistics use the fixed 356-cell mask and 49 rolling 24-hour windows starting at 0–48 h. Event counts use the 50-mm threshold and include overlapping windows; zero denotes no threshold exceedance, not necessarily no rain.

Driver	Mean CSI	$n$	Pooled Bias	24 h MSE (mm 2 )
GFS	0.101	70	0.996	128.57
$\alpha=0$	0.145	73	1.704	240.74
$\alpha=0.25$	0.136	73	1.672	260.23
$\alpha=0.50$	0.129	73	1.498	205.56
$\alpha=0.75$	0.122	73	1.284	181.77
$\alpha=1$	0.121	73	2.201	257.70

Appendix

Table 5: Supplementary starts 0–72 h, requiring rain forecasts through 96 h.

n

counts defined pooled CSI windows out of 73. Bias retains all false alarms, including in windows without reference events.

Figure 4: 20 July 2023 initialization, accumulation from +25 to +49 h. The window is chosen from reference precipitation only. Fields are conservatively mapped to the verification grid; red contours mark 50 mm. The same discrete color scale is used for all drivers and all three map figures.

Figure 5: 30 July 2023 initialization, accumulation from +0 to +24 h, selected by the same reference-only rule. Fields use the common verification grid, color scale, and 50-mm contour convention.

Figure 6: 10 August 2023 initialization, accumulation from +30 to +54 h, selected by the same reference-only rule. All six drivers are displayed on the common verification grid and color scale.

Explore similar work

Does Aurora Encode Atmospheric Structure? Latent Regime Analysis and Attribution
Jun 24, 2026Emma Kasteleyn, Ana LucicWeather Forecasting
University of Amsterdam
WeatherBridge: Forecast-Aware Multi-Source Input Adaptation for Pretrained Weather Forecasting Models
Oct 10, 2026Mutian Xi, Xinyu Su, Xiaohui Zhong +1Weather Forecasting Multi-Source Domain Adaptation
Artificial Intelligence Innovation and Incubation Institute, Fudan University · Shanghai Academy of AI for Science · FuXi Intelligent Computing Technology Co., Ltd.
Global Evaluation of AI and NWP Precipitation Forecasts During Atmospheric River Events
Sep 26, 2026Marina Vicens-Miquel, Taylor Mandelbaum, Amy McGovern +2Precipitation Forecasting Weather Forecasting
School of Meteorology and School of Computer Science, University of Oklahoma, Norman, 73072, OK, USA. · NSF AI Institute for Research on Trustworthy AI in Weather, Climate, and Coastal Oceanography, Norman, 73072, OK, USA. · Brightband, San Francisco, CA, USA.

Jun 24, 2026J/K move · Enter open · S save

Does Aurora Encode Atmospheric Structure? Latent Regime Analysis and Attribution

Emma Kasteleyn, Ana Lucic

University of Amsterdam

Weather Forecasting

Open paper AskPDF arXiv

ML foundation models are able to emulate atmospheric dynamics accurately and efficiently but operate as opaque ``black boxes''. We investigate the internal representations of the Aurora model using spatially pooled PCA and layer-wise relevance propagation (LRP). We find evidence that Aurora's latent space is primarily organized by seasonal cycles, whereas extreme storm events do not form a linearly separable cluster. LRP indicates that the model attends to features consistent with the 3D vertical structure of the Great Storm of 1987. Perturbation tests show masking relevant regions degrades forecasts

3.31\times

more than random masking. These findings suggest that Aurora learns meteorological coherence and vertical structure without explicit instruction.

Abstract

Explore similar work

Does Aurora Encode Atmospheric Structure? Latent Regime Analysis and Attribution

WeatherBridge: Forecast-Aware Multi-Source Input Adaptation for Pretrained Weather Forecasting Models

Global Evaluation of AI and NWP Precipitation Forecasts During Atmospheric River Events