Summary of Key Points
Recently, the AI research agent MIRA from the Deep Principles team made a significant breakthrough: through recursive self-training, it independently completed the entire process from code reconstruction, data cleaning to training strategy design, resulting in the creation of the Material Property Assessment (MPA) model. This model has surpassed all previous global best performances (SOTA) across 40 experimental prediction tasks, with an average error reduction of 10% and a maximum decrease of 51%. More importantly, this marks the official initiation of the “flywheel” of AI self-evolution in the field of materials science, potentially accelerating the arrival of General Artificial Intelligence (AGI) than we anticipated.
Detailed Explanation
#### 1. Traditional Material Models Rely on Brute Force, MPA Uses Ingenious Approaches to Overcome Bottlenecks
The previous Suiren model from Shanghai was a typical example of “brute force aesthetics”: it used 320 high-end GPUs and 70 million data points to create a model with 1.8 billion parameters, winning the rankings at the time. However, it had a fatal flaw—it could only predict “computational properties” (such as those that can be calculated by quantum chemistry software) but couldn’t handle the critical “experimental properties” encountered in actual research and development (such as boiling point, toxicity, solubility). Why? Experimental properties are extremely challenging to predict due to limited data (each experiment takes several days), high noise (various results from different laboratories), and fundamentally different physical principles behind each property (for example, boiling point is related to intermolecular forces, while toxicity is related to biological mechanisms). Simply accumulating more data and hardware doesn’t solve these issues of “physical diversity.” MIRA targets this pain point by using AI to find innovative methods instead of brute force.
#### 2. AI as a “Full-Stack Researcher”: Thinking and Coding Independently
MIRA is not just an ordinary research tool; it acts like an all-around research assistant:
- Independent Thinking: The team poses a problem to MIRA: “Given 3D molecular structures and experimental labels, how do we design a model for predicting multiple properties?” MIRA systematically analyzes all possible approaches and ultimately chooses the UniMol 3D framework as the basis.
- Independent Code Modification: It directly rewrites the source code of the existing model (not just adjusting parameters). For example, it identifies redundant modules, redesigns the data flow, and standardizes the interfaces for pre-training, intermediate training, and post-training. Throughout this process, humans only need to pose questions and confirm the direction; they don’t have to write a single line of code. This is the biggest difference between MIRA and conventional tools—it can manipulate the underlying code that defines the model architecture and training pipeline, not just tweak surface-level hyperparameters.
#### 3. AI Possesses “Research Intuition”: Automatically Cleansing Data and Understanding Physics
Experimental data comes from various databases and is often in disarray (inconsistent units, duplicate samples, incorrect labels). MIRA can not only automatically handle these basic issues but also uses “physical intuition” to assess the validity of the data. For instance, if the boiling point data for a molecule doesn’t match its molecular weight or functional group composition, MIRA will automatically discard that data. This task would normally take weeks of manual review by experts.
#### 4. Three-Stage Training: Combining LLM Experience with Physical Principles to Make the Model Smarter
The core of MIPA is the “three-stage training framework” designed by MIRA, which integrates the training methods of large language models (LLMs) with material physics:
- Pre-training: Learns the general 3D structures of 64 million molecules to lay a foundation.
- Physically-aligned Intermediate Training: A critical step where only content sharing the same physical mechanisms with the target properties is learned (for example, thermodynamic properties are additive, so MIRA focuses on developing additive capabilities).
- Post-training: Uses the Huber loss function to reduce the impact of outliers (more stable than the traditional MSE loss) and designs a “mixed approach” that handles properties varying with molecule size (such as heat of combustion) and those that remain constant (such as flash point). These improvements enable the model to automatically adapt to different physical principles, significantly reducing errors (e.g., a 51% reduction in heat of combustion errors).
#### 5. AI Self-Improves, and the AGI Flywheel Begins to Spin
MPA’s performance is impressive: it has achieved SOTA across all 40 tasks and performs more reliably on unseen molecules (with only a 6% decrease in performance degradation compared to Suiren). However, the real significance lies in its implications:
This represents the most convincing implementation of “AI for AI”—MIRA uses AI to rewrite model code, optimize training data, and design training strategies, ultimately creating a stronger AI model. The role of humans has shifted from “executors” to “goal setters.” Once this cycle of recursive self-improvement begins, each iteration will be faster than the previous one (as a more advanced AI can improve the next version even quicker). From automatically writing code to conducting research and then self-improving, the capabilities of AI are expanding rapidly, and AGI may be closer than we think.
In One Sentence
AI is now capable of acting as a researcher, improving its own models. Materials science is just the beginning—the gears of AGI have already started turning, and the future might arrive sooner than we anticipate.