Explore the HUAWEI CLOUD Pangu Model: AI for Industries

The big model is the core of the new round of AI development, which has shown great potential in promoting industrial intelligent upgrading and will form a trend in the next three years.

In 2021, HUAWEI CLOUD officially released the Pangu Basic Big Model, including the CV Computer Vision Big Model, the NLP Natural Language Processing Big Model, and the Scientific Computing Big Model. On top of the basic big model, HUAWEI CLOUD practices AI for industries and has successively launched Pangu industry big models such as mining, drug molecules, electric power, meteorology, and ocean waves, accelerating the digitalization process of all walks of life.

Today, HUAWEI CLOUD AI has more than 1,000 projects in various industries, and based on the in-depth understanding of the industry, the HUAWEI CLOUD Pangu model can be better implemented in core business scenarios in the industry. This special program of Pangu Large Model will take you to explore the industry application of Pangu Large Model.

Video loading...

▲ Pangu model special program of "Exploring the Secret Cloud New Knowledge"

Intelligent detection of freight trains, Pangu track large model escorts railway logistics safety

With the economic recovery at home and abroad, the frequency and load of freight railways have ushered in a new climax.

In the traditional freight railway inspection, TFDS (Dynamic Image Detection of Freight Car Running Fault) system, as an important part of 5T detection technology, dynamically collects images of the bottom accessories and side parts of the train body through the "electronic eye" of high-speed shooting, and determines whether the freight train has a fault in real time.

In a 50-car truck, the "electronic eye" will take 4,000 photos, and the inspector will review one picture per second. Dynamic inspectors not only work intensively and difficultly, but also need to have a high level of professional theory and practical application of vehicles, and must complete the fault analysis of the whole train in a short time to ensure the operation safety of the whole train.

Based on the existing equipment and platform architecture, HUAWEI CLOUD has launched the TFDS intelligent fault identification solution based on the Pangu Rail Industry Big Model, which implements global analysis from image acquisition, data transmission and reception, train splitting, and real-time fault identification, and can accurately predict global faults such as multiple stations and vehicle associations.

The Pangu Railway Industry Model has five core competencies:

• Self-supervised industry pre-trained models

Based on the comparative representation learning method of semantic similarity samples and hierarchical semantic aggregation, Pangu pre-trained large model uses million-level unlabeled railway industry images to generate a large model of rail industry.

• Image quality is automatically enhanced & evaluated

The enhanced image is automatically evaluated through the underlying visual features and high-level visual features, and the normal image is further identified by fault, and the abnormal image is returned for manual review.

• Rely on model priori template matching

Establish the relative position template of parts based on known model information, and interpretably predict abnormal parts such as falling off, loss, misalignment, etc.;

• Fault location and identification of small samples

Based on the pre-trained large model of the rail industry, combined with the current optimal object detection and image recognition framework, the component positioning and fault recognition are carried out, and the generalization ability is stronger, and only 1/3 of the traditional sample can be completed.

In practical applications, the single photo recognition of Pangu track large model only takes 4 milliseconds, which can intelligently filter 95% of normal pictures, realize the automatic identification of more than 400 kinds of faults and the "zero false alarm" of serious faults, which is more accurate than manual identification, greatly improves the operation efficiency of TFDS system, and dynamic inspectors can free up more energy to deal with more difficult map identification work to ensure the safe operation of trains.

AI-assisted drug design, Pangu drug molecule large model accelerates new drug research and development

Since the discovery of daptomycin in 1987, no new antibiotics have been developed in nearly 40 years. It takes drug discovery experts more than 10 years and more than $1 billion to develop a new drug.

To help drug R&D experts efficiently select small molecules suitable for over-the-counter drugs from a large number of drug molecules, HUAWEI CLOUD and the Shanghai Institute of Materia Medica, Chinese Academy of Sciences launched the Pangea Drug Molecular Model, which is based on the capabilities of full-process AI-assisted drug design, using target prediction, molecular design, activity evaluation, and toxicity screening as the starting point, helping pharmaceutical companies achieve rapid, accurate, and low-cost drug discovery, opening up a new model of drug development.

• In terms of virtual screening of drugs

Relying on HUAWEI CLOUD's innovative iFitDock algorithm and virtual screening service, the druggability prediction accuracy of the Pangu drug molecule model is 20% higher than that of traditional methods, thereby improving the efficiency of drug screening by ten times.

• In terms of drug optimization

Based on the structure optimizer of HUAWEI CLOUD Pangu drug molecule model, R&D experts can optimize the lead drug in a targeted manner, and reduce the possible toxic side effects on normal cells through more scientific drug structure design.

Four core technical characteristics of Pangu drug molecule model:

• "Graph-Sequence Asymmetric Conditional Variational Autoencoder"

The new "graph-sequence asymmetric conditional variational autoencoder" deep learning architecture is proposed to better extract the key molecular feature fingerprints of compounds and improve the accuracy of downstream tasks.

• Very large-scale compound characterization model training

The chemical structure of 1.7 billion small molecules was pre-trained, and the structural reconstruction rate and uniqueness were superior to existing methods.

• Generate a library of 100 million innovative drug-like small molecules

Its structural novelty is 99.68%, creating the possibility of discovering new drugs;

• Achieve leading-edge drug discovery mission performance

It achieves optimal performance in terms of compound-target interaction prediction, compound ADME/T attribute score, compound molecule generation and optimization, etc., enabling the whole chain of drug discovery.

Professor Liu Bing of the First Affiliated Hospital of Xi'an Jiaotong University, with the assistance of the Pangu drug molecule model, has developed a super antibacterial drug Drug X, which is expected to become the world's first new target and new class of antibiotics in the past 40 years. HUAWEI CLOUD Pangu Drug Molecule Model shortens the R&D cycle of lead drugs from years to months, reducing R&D costs by 70%. The combination and innovation of AI technology and basic science not only solves the pain points of high R&D cost and long time cycle, but also provides a stage for start-up scientific research teams to display their capabilities.

Make the wind and clouds measurable, and the Pangu meteorological model accurately presents the typhoon trajectory

In the meteorological climate prediction task, in addition to short-term weather forecasting, global medium and long-term forecasting is also the industry's most concerned and important forecasting task, which aims to predict the state of the atmospheric system in the next 14 days and plays a pivotal role in meteorology, navigation, agriculture, tourism and other industries.

Although artificial intelligence technology has been widely used in the field of meteorological prediction, due to the complexity of physical processes in atmospheric systems and the huge scale of resources required to solve atmospheric models, medium- and long-term weather forecasts based on traditional numerical methods usually accumulate errors, resulting in low accuracy and hours of computation on supercomputers.

Based on nearly 40 years of global meteorological data, HUAWEI CLOUD's Pangea Meteorological Model surpasses the current strongest numerical forecasting method (the IFS system of the European Meteorological Center) in medium- and long-term deterministic forecasting, and is the industry's first global AI weather prediction model with accuracy that exceeds traditional numerical forecasting methods. The average forecast error is reduced by 10%-15%, and the speed is increased by more than 10,000 times, realizing second-level global weather forecasting.

Core technical features of Pangea Meteorological Model:

• 3D high-resolution neural network

The first use of 3D Earth-Specific Transformer: Compared with two-dimensional neural networks and low-resolution neural networks, the horizontal spatial resolution of Pangea Meteorological Large Model reaches 0.25∘×0.25∘, about 28 km x 28 km, which can accurately predict fine-grained meteorological features. In the time dimension, the Pangea Meteorological Model increases the prediction frequency from 6 hours/time to 1 hour/time, making the weather prediction results more accurate.

• Hierarchical time-domain aggregation policies

Hierarchical time domain aggregation strategy is used: four models with different forecast intervals (1-hour interval, 3-hour interval, 6-hour interval, and 24-hour interval) are trained, so that the number of iterations to predict meteorological conditions at a specific time is minimized, thereby reducing the iteration error and avoiding the training resource consumption caused by recursive training.

HUAWEI CLOUD's Pangea Meteorological Model has demonstrated the advantages of accuracy and speed in forecasting extreme weather processes such as typhoons:

• In August 2022, the Pangu Meteorological Model achieved a second-level prediction of the trajectory and landing time of Typhoon Saddle, with an accuracy rate of 90%, far exceeding the industry average.
• From May 22 to 23 this year, this year's No. 2 typhoon "Mawa" quickly strengthened from 38 m/s (typhoon level) to 60 m/s (super typhoon level) in 24 hours.

According to the Central Meteorological Administration, HUAWEI's cloud Pangu model performed well in the path prediction of Mawa, predicting that it would turn its path in the eastern waters of Taiwan Island five days in advance.

The industrial transformation triggered by artificial intelligence is changing every industry, and artificial intelligence is also playing an important role in more and more industry scenarios. HUAWEI CLOUD focuses on AI for industries to improve the general capabilities of large models, meet the real-world needs of customer service scenarios, standardize, replicate, and mass produce AI development, accelerate the penetration of AI into thousands of industries, and promote human society into the intelligent world. HUAWEI CLOUD Developer Conference 2023 (Cloud) Conference will kick off in Dongguan on July 7, and HUAWEI CLOUD Pangu model will usher in a major upgrade, so stay tuned!

Follow @HUAWEI CLOUD to learn more