When we talk about AI, we often think of chatty robots or models that can “paint.” However, behind these visible interfaces, a more substantial transformation is taking shape: AI is turning the “discovery of materials” into a searchable, designable, and verifiable engineering process, directly targeting some of humanity’s most challenging issues such as clean energy, carbon capture, and water purification.

The following text distills this transformation into five key insights using the approach of “first ask if, then ask why.”

1｜Behind Simple Metaphors Lies a Process-Level Revolution

Is it: Is AI turning “finding materials” into “searching for materials”? Yes. Why: Because the objects of “search” are no longer just materials in known literature, but all possible molecules and crystal structures that can be created.

For example, on the CuspAI platform—it’s like a material search engine for users: input target properties (such as “porous materials that efficiently adsorb CO₂ under specific conditions”), and the system completes the

generation-evaluation-optimization-synthesis feasibility judgment end-to-end process.

• In the past, the process relied heavily on computational-experimental trial and error, which was lengthy;

• Now, with generative AI embedded with physical priors and high-throughput evaluation at its core, the “trial and error” is moved to the model space;

• It claims to increase the speed of discovery to ten times the original level, compressing “years” into “months.”

“Imagine a search engine that not only finds existing materials but can also search for potentially creatable materials. Our AI can generate and evaluate on demand.” — Max Welling

Key points for practitioners: Integrate “problem-property-structure-process-experimental validation” into a streamlined process rather than fragmented steps.

2｜Bigger is Not Always Better: Counterintuitive Lessons from Cutting-Edge AI

Is it: Is a large model + massive data always optimal in scientific scenarios? Not necessarily. Why: Scientific exploration emphasizes explainability, retrainability, and iterability. NIST research shows that high-density information small samples can train smaller, faster, and more interpretable models that sometimes achieve performance comparable to larger models; and the retraining cost is lower, with more flexible iterations.

• Small models (like random forests) can be incrementally retrained in < 30 seconds;

• Large graph neural networks take > 12 hours;

• In many “learn-by-doing” experimental-computational loops, iteration speed surpasses single-point extreme accuracy.

“Bigger is not necessarily better; training smaller, faster, and interpretable models with carefully selected data, accepting some performance compromise, leads to greater scientific insights.” — NIST Research Perspective

Key points for practitioners: Choose models centered around the problem, prioritizing “iterable closed loops” over blindly chasing parameter scale.

3｜The Key is Not Just in “Compounds” but in “Usable Materials”

Is it: Does AI generating millions of “new compounds” mean a breakthrough in materials? No. Why: Compounds (compound) ≠ Usable Materials (material). Real materials need to be stable, synthesizable, and scalable, achieving optimal overall performance under target conditions.

• Some AI-generated structures may not be novel or may not be producible in the real world;

• The effective path is to incorporate laboratory testing and process constraints into the AI design closed loop: design-synthesis path-experimental validation-parameter feedback.

This is akin to the classic joke of the “spherical cow”: Simplified models (spherical cows) help us quickly identify feasible directions, while high-fidelity models and experiments bring solutions to reality.

Key points for practitioners: Elevate “synthesizability, scalability, cost, and regulations” to the same level as performance in the objective function, and perform multi-objective optimization.

4｜AI’s “Ultimate Challenge” Aligns with Humanity’s “Ultimate Challenge”

Is it: Are breakthroughs in materials aimed at societal-level challenges starting to enter a speed phase? Yes. Why: AI is actively addressing the following hard problem scenarios:

• Climate Change: Designing new carbon capture materials to reduce CO₂ removal costs;

• Clean Water Sources: Collaborating with chemical companies (like Kemira) to explore materials for removing PFAS (“forever chemicals”);

• Clean Energy: Working with automotive and energy companies to advance sustainable energy materials.

“The next decade will present many challenges, some triggered by AI and some solved by AI. Accelerating new material design with AI to address climate change is an impressive mission.” — Geoffrey Hinton

Key points for practitioners: Focus on topics around “performance/cost inflection points” and “barriers to large-scale deployment,” using AI to shift technical risks forward and move failure costs forward.

5｜A “Dream Team” is Not Just a Stack of Resumes, but Turning Mission into a Collaborative “Common Language”

Is it: Does a top team = top algorithms + top chemistry + industrial implementation? Yes. Why: The success of materials AI relies on deep interdisciplinary integration and a shared mission:

• For example, at CuspAI: Co-founder Max Welling (co-author of VAE, over 180,000 academic citations) × Chad Edwards (quantum chemistry/industrialization), supported by Hinton, LeCun, and ecosystem partners;

• Aligning top AI expertise with chemistry/process knowledge is essential to transform theoretical breakthroughs into produced solutions.

“We put the best chemists and the strongest machine learning engineers together for seamless collaboration.” — Max Welling

Key points for practitioners: Clearly define a common “problem statement” and “acceptance criteria” to align different disciplines on the same dashboard of metrics.

Action Checklist: Turning “Hype” into “Pathways”

1. First ask if: In your scenario, is the biggest bottleneck performance, cost, compliance, or scalability?

2. Then ask why: Break down the bottleneck into quantifiable objective functions (such as selectivity/energy consumption/cycle stability/material cost/synthesis steps).

3. Build a closed loop:

• Generative material design (including physical constraints)

• Synthesizability and stability assessment (including retrosynthesis/thermodynamics/kinetics)

• Experimental micro-flux validation + data feedback

Select models rather than idolizing models: Prioritize a combination of “fast iteration, explainability, and retrainability”; introduce larger models for fine-tuning at critical windows.

Establish verifiable milestones: Use public benchmarks and third-party experiments to calibrate the real extent of “speeding up/cost reduction.”

Organize collaboration: Align chemistry, materials, processes, and data science on the same dashboard to achieve common goals and manage risks.

Conclusion: The Future Can Be Not Only Discovered but Also “Designed”

AI is advancing materials science from “retrieving the known” to “searching for possibilities,” from “trial and error” to “goal-driven generation and validation.” The blueprint for a sustainable future no longer needs to be buried in a decade of experiments; it can be defined as an objective function, written into models and processes, and then validated as products and industries.

If you are in a related field, consider asking yourself: Can your problem be rewritten as an optimizable objective? Can your process be rewritten as a closed loop? When the answer is “yes,” the search box for materials truly opens.