Machine Learning Bias Correction Distorts Warming Signals in Climate Projections

A new evaluation of climate model correction techniques has found that the methods most skilled at reproducing past weather are not necessarily the ones that preserve future warming signals, a distinction with direct consequences for regional climate projections. The study, published in Theoretical and Applied Climatology, focused on Jiangsu Province in eastern China and systematically compared statistical and machine learning approaches for correcting systematic regional biases in CMIP6 climate models.

Climate models routinely show persistent errors when simulating regional climate. The research team built an evaluation framework integrating both classical statistical methods and machine learning approaches, testing twelve correction techniques in total: seven statistical and five machine learning based. Observations from the CN05.1 dataset served as the reference benchmark throughout.

The first stage of the analysis identified the best-performing model among twelve CMIP6 climate models using Taylor diagrams, a standard tool for comparing correlation, variance and error patterns against observational data. ACCESS-CM2 emerged as the strongest performer for the region. The twelve correction methods were then evaluated with an independent temporal hold-out design, in which the methods were trained on observations from 1961 to 2000 and validated on the separate 2001 to 2014 period, a design intended to test genuine out-of-sample skill rather than fit to the training data.

At the monthly timescale, a convolutional long short-term memory network, known as CNN-LSTM, achieved the highest correlations with observations for all three climate variables examined and produced the lowest error for precipitation in this case study.

However, the study also exposed a methodological weakness in how correction methods are commonly ranked. The researchers constructed a composite performance index to summarize method skill, and found that the resulting rankings depended on how the index was built. No weighting scheme agreed consistently with independent benchmarking, meaning that summary scores constructed by analysts can lead to different conclusions about which method is best, depending on choices that are not dictated by the data themselves.

The most consequential finding concerned the preservation of climate change signals. At the daily scale, Random Forest, carried forward as a representative tree-ensemble method, performed best among the three sources compared on error-based criteria. Yet when the researchers examined how well corrected output retained the projected warming trend, Random Forest preserved less than five percent of it. Quantile-based methods, by contrast, retained more than ninety-nine percent of the projected trend. Random Forest also reproduced only three of thirteen extreme climate indices within ten percent of observations.

This divergence demonstrates that historical skill and change-signal preservation are distinct properties of a bias correction method. A technique that minimizes error against past observations can simultaneously suppress or distort the physical trend that the underlying climate model projects for the future. Because the purpose of bias correction in projection studies is to adjust model output without erasing the model’s climate response to emissions scenarios, the study’s authors treated signal preservation as a decisive criterion.

For the final projections, the study therefore used quantile delta mapping, a quantile-based correction approach, applied to the best-performing model. Under the high-emissions SSP3-7.0 scenario, the projections indicate that tropical nights, defined as nights with minimum temperatures above a threshold, would increase by roughly seven days per decade, while warm days would increase by roughly ten days per decade across Jiangsu Province. Annual total precipitation, in contrast, showed no significant trend under the same scenario. Projections were also analyzed under the lower-emissions SSP1-2.6 pathway.

The authors are explicit about the scope of these results. The projections derive from a single bias-corrected model under two scenarios rather than a probabilistic assessment spanning many models, so the reported trends should be read as a case study of one region and one model rather than a full uncertainty range. Even so, the central methodological conclusion travels beyond Jiangsu: evaluations of bias correction that rely solely on historical error metrics can favor methods, including popular machine learning ensembles, that fail to carry the warming signal forward, and projection studies should benchmark signal preservation independently before selecting a correction technique.

Subject of Research: Evaluation of statistical and machine learning bias correction methods for CMIP6 climate model projections of extreme climate events in Jiangsu Province, China

Article Title: Comprehensive evaluation of multi-method bias correction for cmip6 climate models and future scenario simulation of extreme climate events: a case study of Jiangsu Province, China

Article References: Ruan, P., Chen, Y., & Feng, Z. (2026). Comprehensive evaluation of multi-method bias correction for cmip6 climate models and future scenario simulation of extreme climate events: a case study of Jiangsu Province, China. Theoretical and Applied Climatology, 157(11), Article 755. https://doi.org/10.1007/s00704-026-06659-0

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06659-0

Keywords: CMIP6, bias correction, machine learning, climate extremes, Jiangsu Province, quantile delta mapping, SSP scenarios, climate models, extreme heat, precipitation, ACCESS-CM2, CNN-LSTM