AI Decodes Historical Chinese Archives to Reveal Lost Climate History

Sida He, Lingxi Xie, Xiaopeng Zhang, Qi Tian

Abstract

Historical archives contain qualitative descriptions of climate events, yet converting these into quantitative records has remained a fundamental challenge. Here we introduce a paradigm shift: a generative AI framework that inverts the logic of historical chroniclers by inferring the quantitative climate patterns associated with documented events. Applied to historical Chinese archives, it produces the sub-annual precipitation reconstruction for southeastern China over the period 1368-1911 AD. Our reconstruction not only quantifies iconic extremes like the Ming Dynasty's Great Drought but also, crucially, maps the full spatial and seasonal structure of El Niño influence on precipitation in this region over five centuries, revealing dynamics inaccessible in shorter modern records. Our methodology and high-resolution climate dataset are directly applicable to climate science and have broader implications for the historical and social sciences.

Introduction

To understand the trajectory of modern climate change, researchers require long-term baselines. However, high-quality instrumental records—derived from satellites and weather stations—rarely extend back more than a century. For the centuries prior, science relies on physical proxies like tree rings or ice cores. While valuable, these proxies often lack sub-annual (seasonal or monthly) resolution.

Conversely, historical archives, such as China’s vast imperial records, provide high-resolution descriptions of weather events. However, these records are qualitative; a chronicler might note a "severe drought" without providing a rainfall measurement in millimeters. The challenge lies in "inverting" this logic: transforming sparse, descriptive text into a continuous, quantitative map of precipitation across an entire region.

Methodology: The 3D Diffusion Model

Researchers have addressed this challenge by developing a generative AI framework based on a 3D Diffusion Model (utilizing a U-Net architecture). This approach moves beyond traditional linear regression, which often struggles with sparse historical data, to learn the complex, joint probability distribution of climate fields.

The generative AI framework for sub-annual climate reconstruction, contrasting previous qualitative methods with the new quantitative diffusion model approach.

The model operates on several key principles:

  1. Inverse Logic: The model is trained to infer the underlying quantitative climate patterns (precipitation z-scores) that most likely gave rise to the specific sequence of events documented by historical chroniclers.
  2. Synthetic Training: To overcome the lack of paired "text-to-numbers" data from the pre-instrumental era, the team used modern climate model simulations (CESM1-CAM5) to create "synthetic" history. Simulated rainfall was translated into discrete "events" to teach the AI the relationship between the two.
  3. Spatiotemporal Consistency: The model treats two spatial dimensions and the 12 months of the year jointly, ensuring that reconstructed rainfall for a given month is physically consistent with both the preceding months and the surrounding geography.
  4. Calendar Alignment: A "periodic conditioning prompt" was implemented to synchronize the shifting timing of the Chinese lunar calendar with the Gregorian calendar, ensuring historical accuracy.

Results: Quantifying Five Centuries of Precipitation

The framework produced the first-ever sub-annual precipitation reconstruction for southeastern China spanning 544 years (1368–1911 AD). On validation datasets, the model demonstrated strong predictive skill, achieving an annual precipitation R-squared (RSQ) of 0.61 and a Coefficient of Efficiency (CE) of 0.70.

Reconstructed annual and seasonal precipitation anomalies for the years 1593, 1640, and 1877, highlighting the severe 1640 drought that contributed to the fall of the Ming Dynasty.

A major achievement of the study is the quantification of iconic historical catastrophes. For instance, the 1640 Chongzhen Drought, a pivotal factor in the collapse of the Ming Dynasty, was quantified with annual precipitation z-scores below -2 across much of northern China. The model also captured the "canonical dipole pattern" of El Niño—typically characterized by wet conditions in the southeast and dry conditions in the north—and revealed a seasonal phase reversal that aligns with modern observations.

Discussion and Future Outlook

The study reveals that the influence of the El Niño-Southern Oscillation (ENSO) on Chinese precipitation has evolved over time. While 19th-century patterns resemble modern records, earlier centuries show distinct variations, providing a new frontier for studying long-term climate oscillations.

While the current model uses binary inputs (event vs. no event), future iterations could incorporate intensity gradations and document survival probabilities to further refine accuracy. The ultimate goal is a multi-modal AI that fuses textual records with physical proxies to create a definitive historical climate map. This work demonstrates that generative AI can unlock centuries of "lost" climate data hidden within qualitative historical archives.

Summary updated

Abstract

Historical archives contain qualitative descriptions of climate events, yet converting these into quantitative records has remained a fundamental challenge. Here we introduce a paradigm shift: a generative AI framework that inverts the logic of historical chroniclers by inferring the quantitative climate patterns associated with documented events. Applied to historical Chinese archives, it produces the sub-annual precipitation reconstruction for southeastern China over the period 1368-1911 AD. Our reconstruction not only quantifies iconic extremes like the Ming Dynasty's Great Drought but also, crucially, maps the full spatial and seasonal structure of El Ni$ñ$o influence on precipitation in this region over five centuries, revealing dynamics inaccessible in shorter modern records. Our methodology and high-resolution climate dataset are directly applicable to climate science and have broader implications for the historical and social sciences.
BibTeX
@misc{he2026ai,
  author = {Sida He and Lingxi Xie and Xiaopeng Zhang and Qi Tian},
  title = {{AI Decodes Historical Chinese Archives to Reveal Lost Climate History}},
  year = {2026},
  eprint = {2601.22458},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.22458},
}
Figure