Deep Learning
+ First Principle
R&D Is Still Experience-driven and Experiment-based
Long Duration
1-2 experiments per day 10 years from lab to production
High Cost
1 lab-scale success with a new product can cost $2-20 million
Long Duration
55% experimental data is unstructured
50% experiments are unrepeatable
70% of researchers struggle to repeat results
The ECML Framework for Materials R&D
In mainstream R&D paradigms, Three Gears drive scientific discovery in their own way.
Large Gear
Wet Lab Experiments
Seconds | Hundreds of Dollar
Despite offering the highest accuracy that closely mirrors real-world conditions, this approach is the slowest. It typically takes months or even years to complete a series of experiments—including material synthesis, performance testing, and structural characterization—to achieve a closed loop for material R&D verification. This also results in exorbitant R&D costs, ranging from millions to tens of millions.
Accuracy
Medium Gear
High-Throughput Computing
Days to Weeks | Hundreds of Thousands of Dollar
Featuring relatively fast processing speeds and high accuracy, this approach includes multiple hierarchical simulation methods: DFT calculations (revealing reaction mechanisms at the electronic scale), Molecular Dynamics (MD) (nanosecond-scale atomic motion simulation), and Kinetic Monte Carlo (KMC) (predicting long-range diffusion processes). It typically requires days or weeks, with costs on the order of hundreds of thousands. However, it is constrained by the inherent trade-off between computational accuracy and spatiotemporal scale limitations.
Accuracy
Small Gear
Machine Learning
Months to Years | Millions of Millions Dollar
Leveraging technologies such as Molecular Fingerprints (material classification in high-dimensional feature spaces), Machine Learning Force Fields (MLFF) (replacing DFT to enable nanosecond molecular dynamics), Graph Neural Networks (GNN) (constructing micro-macro structure-property relationships), and Generative AI (directly generating molecular structures and chemical reactions based on diffusion models), this approach can complete performance prediction and structure generation within seconds. However, its generalization capability is bottlenecked by the scarcity of material data.
AI Decision-Making
R&D Time
Funding/Investment
Accuracy of Results
