In 2024, Muhammad Irfan Virk first visited the China Meteorological Administration (CMA) in Beijing and was impressed by the level of technological progress, noting that the CMA forecasting offices seemed 'almost science fiction.'
Virk was part of a team of meteorologists from the Pakistan Meteorological Department (PMD) sent to the CMA to adapt the weather forecasting platform for Pakistan's needs. This system is known as MAZU, which stands for 'Multi-hazard, Notification, Zero-gap, and Universality.' The name also refers to the traditional Chinese sea goddess who protects sailors.
Zahir Ahmed Babar, Chief Meteorologist of PMD and Acting Director General, emphasized that the AI system cannot stop extreme weather events. However, by providing earlier warnings, it allows forecasters to make more informed decisions. He added that using this information to reduce losses remains at the discretion of the users.
According to Yin Xinxin, scientific attaché of the Chinese Embassy in Islamabad, the MAZU version adapted for Pakistan has been used by PMD since October 2025, in addition to other early warning tools, including satellite data and other weather models.
Yin Xinxin noted that because Pakistan is in urgent need of early warning capabilities and disaster risk reduction, both sides jointly developed and implemented the MAZU system, making Pakistan the first country to receive this innovative development. Cooperation between CMA and PMD has a long history, and according to Yin, it is carried out under the framework of China's 'South-South cooperation on climate change.'
Following floods that occurred on August 26 in neighboring regions of Nepal and China, MAZU is also used to assess and forecast local weather conditions to organize rescue operations. According to China Daily, Chinese authorities use the system to provide Nepal with daily updates on the development and risk forecasts of newly formed barrier lakes, as well as for hydrological monitoring in emergencies.
Although AI-based weather modeling experts welcome the use of the Chinese model, they warn that it is not a panacea for early warning systems. Furthermore, it is too early to judge its accuracy in Pakistan's local climatic and geographical context.
At an April press conference dedicated to further international cooperation to expand MAZU, CMA administrator Chen Zhenlin stated that early warning systems are an economically efficient way to protect life and property against the threat of extreme weather for food and energy security.
According to Chinese news reports, at the World Artificial Intelligence Conference in Shanghai in 2026, the MAZU booth generated huge interest among diplomats, engineers, and journalists. The system's popularity may have increased because Chinese President Xi Jinping mentioned it in his keynote speech at the conference. He announced that over the next five years, China will provide 5000 opportunities for developing countries for AI training and workshops and will allow 30 countries to use the AI-based meteorological warning system, MAZU.
Babar from PMD noted that MAZU can be adapted to the geography, infrastructure, and disaster risks of a specific country by integrating local meteorological data. Pakistani researchers and forecasters have been trained to use the version calibrated for Pakistani data.
The MAZU system has also been implemented in Djibouti, Ethiopia, Mongolia, Jordan, Sri Lanka, and the Solomon Islands. China also supports over 40 countries through cloud virtual trials of the system.
Asim Javed, CEO of AI Geo Navigators, which uses AI, mapping data, and remote sensing to assess weather and environmental risks, finds it valuable that the platform integrates traditional meteorology, satellite and radar data, ground station observations, and local calibration with AI analytics and warnings. He emphasized that integrating all this input data in one place changes forecasters' approach to work.
Climatologist Imran Khalid warns that AI-based weather tools support early warning systems but are not 'autonomous systems sufficient to fill our gaps... at least for now.' Babar explained that an effective early warning system requires a comprehensive government approach based on four components: disaster risk management, detection and forecasting, dissemination and communication, and preparedness and response.
This holistic approach is supported by Junaid Yamin, co-founder of the Pakistani weather forecasting service WeatherWalay. He also stated that the success of the hydrometeorology sector ultimately depends on effective public-private partnerships. Yamin stressed that no single organization can tackle climate change alone; it requires overcoming barriers, sharing data, and aligning the efforts of government, business, and civil society around a common mission.
Javed agreed with this, asserting that meteorological data should be 'open source.' He explained that rain and heat trigger numerous weather hazards affecting business—from telecommunications and banking to agriculture. Access to weather data can help businesses develop products that mitigate disaster damage.
The World Meteorological Organization (WMO) promotes freer exchange of observational meteorological data within its unified 2021 data policy and 2024–27 strategy, which emphasizes partnership with the private sector.
Babar explained that the atmosphere is 'inherently chaotic,' and even small changes in air pressure, temperature, wind, and humidity can make accurate forecasting impossible using only traditional methods. This is where systems like MAZU can help, as they can solve complex mathematical equations in seconds, allowing forecasters to make faster and more timely decisions.
However, as Khalid noted, these systems are only as good as the weather data fed into them. Without quality data, forecasts will just be guesses. Such programs can predict large-scale events, such as monsoons and heatwaves. Nevertheless, precise forecasts—such as when a monsoon will reach a specific valley in a mountainous region or whether heat will cause glacial lake floods—require dense, uniformly distributed radar coverage, which Pakistan lacks.
Glacial lakes formed by melting mountain glaciers pose a growing threat in the Himalayan watershed. These floods can be caused by earthquakes, landslides, or excessive accumulation of meltwater. Javed noted that Pakistan presents a complex case: ground stations are sparse, and in the northern mountainous regions of the country, elevation can change sharply over just a few kilometers. Events leading to loss of life—highly localized heavy rainfall and flash floods—occur on scales that most global models cannot predict. MAZU cannot predict weather at the 1 km level, like most US and European models, which typically operate at a resolution of about 9 km.
Another problem, according to Javed, is the common misconception that AI systems automatically yield highly accurate results. He explained that everything still depends on observations and models. The model learns from the past, but the events for which we need warnings the most often have the least historical record. These gaps limit the accuracy of local weather forecasting. Khalid warned that without detailed ground data, AI tools can 'hallucinate,' producing confident local forecasts without a real observational basis, which could lead to serious consequences.
He also noted that models can miss highly localized events, such as intense rainfall across several blocks or the exact time and location of a cyclone making landfall, leaving communities unprepared. Therefore, localized radar data is critically important.
Model outputs must be independently verified, and the systems themselves must be trained considering Pakistan's realities. This includes training on historical weather patterns and indicating uncertainties in results when data gaps exist. Furthermore, results must account for institutional limitations, such as delays in inter-governmental communication, as well as shortages of expert personnel and monitoring equipment, as noted by Khalid.
Javed also raised the issue of accountability. Using an AI-supported system like MAZU introduces additional model outputs or automated recommendations. These must be documented, while the authorized PMD forecaster retains responsibility for issuing the public warning. He insisted that every warning must have a clear audit trail showing the observations, model outputs, thresholds, and professional judgment behind it. He added that MAZU should make this judgment 'more precise, not optional.'
To verify whether MAZU truly performs better than existing systems, Javed proposed using three simple criteria. First, he would check the accuracy of MAZU forecasts compared to the existing PMD system and an independent benchmark, such as a recognized global forecasting system. Second, a threat detection test would be conducted: 'Can MAZU accurately identify dangerous weather phenomena, or just predict normal weather?' Third, practical value would be assessed: how much extra warning time the system provides and how many false alarms it generates. He concluded that if MAZU proves its superiority in forecasting skill, threat detection, and warning time, it will bring real benefit; otherwise, it will be a well-designed but unverified platform.