Mark Sullivan, our Head of Solution Engineering, has worked with dozens of chemical companies to help them get started with AI. Learn from his experience.
Models Don’t Create Value on Their Own
Even a “perfect” model with flawless predictions generates absolutely zero impact if it isn’t used to make real materials.
Models are tools designed to help you make better materials. They can be developed to uncover insights into the complex relationships between composition, chemical structure, processing conditions, and target technical properties. They’re invaluable for exploring new ideas and running different what-if scenarios across your design space.
But here’s the reality check: no value is actually generated until those ideas are turned into real materials that get tested in a lab environment. No Citrine employee would get on a plane made of AI-predicted but untested materials—and neither would you.
THE BOTTOM LINE
Your customers don’t buy models. They buy adhesives, coatings, and materials that actually work in the real world.
What Improves Model Accuracy?
Unless you work on very simple materials and you already have thousands of data points, you’re unlikely to create a fully predictive model right out of the box.
But here’s the good news: that’s ok.
Your initial model can still suggest experiments in areas where it feels uncertain, allowing it to make better predictions in the future. It can also recommend experiments where it predicts good technical performance, so its hypotheses can be tested. Crucially, it doesn’t need to be 100% accurate to identify uncertainty or determine which experiments are currently most promising.
Over the last 12 years we have seen time and again that new experimental data that is guided by the needs of the moment is dramatically more powerful than historical data because it is directly relevant, contextually understood, and fills gaps that reduce uncertainty. Adding new, targeted data is your best and easiest way to increase your model’s predictive power. The model’s job is to guide you in deciding which experiments to run next to do that.
The Purpose of Modeling is Action
Ask the Right Questions
Which experiment should you run next to learn the most?
Guide Your Exploration
Models identify uncertainty and suggest where to explore your design space
Test Hypotheses
Validate predictions where technical performance looks most promising
Accelerate Learning
New experimental data rapidly increases predictive power and confidence
Perfect Models Don’t Equal Perfect Products
R2 is Not a Product
Even if you could generate a perfect model with 100% accuracy (R² = 1), you can’t sell that statistical achievement to your customer who needs to bring a high-performance adhesive to market.
Insight and predictive power matter only when applied, not when admired in isolation. Your customers buy materials that solve their problems—not models, not algorithms, not impressive correlation coefficients.
The value is realized only when models help you produce adhesives, coatings, or chemicals that actually work in demanding real-world applications.
TIME SPENT MOVING FROM R² =0.6 TO R² =0.65 BY PLAYING WITH MODELS IS TIME WASTED.
Contrary to the instincts of a scientist to get the model perfect before running experiments, business value is achieved faster by using a good enough model and getting to the lab quickly.
What Customers Want
- Materials that meet specifications
- Consistent quality batch-to-batch
- Cost-effective solutions
What Models Provide
- Guidance for experimentation
- Insights into structure-property relationships
- Efficient exploration of design space
- Reduced time to market
The Bridge Between Them
- Strategic experiment design
- Rapid iteration cycles
- Data-informed decision making
- Validated materials in the lab
The Real Power: Inverse Design Workflows
In practice, the transformative power of models emerges when you USE them within an inverse design workflow. This is where theory meets reality, where predictions become prototypes, and where computational power accelerates materials discovery.
01
Define Constraints
Establish your design space boundaries based on real-world manufacturing parameters and allowed raw materials
02
Multi-Objective Optimization
Apply your model to balance competing properties—strength vs. flexibility, cost vs. performance, durability vs. processability
03
Targeted Experiment Selection
Collaborate with your model to identify the most promising formulations that satisfy your constraints
04
Lab Validation
Run focused experiments to test and validate the model’s recommendations
05
Iterate and Refine
Feed results back into the model, improving predictions and guiding the next round of experiments
Apply your model as a tool in a constrained design space with multi-objective optimization to run experiments focused on making the real material your customers need. This workflow doesn’t require a perfect model—it requires a useful model that can intelligently navigate trade-offs and uncertainties.
The model becomes your co-pilot in the innovation journey, helping you avoid dead ends and identify promising directions faster than traditional trial-and-error approaches.
WHY THIS WORKS
Inverse design turns the problem around: instead of predicting properties from compositions, you define desired properties and discover compositions that might achieve them.
The Finish Line is in the Lab, Not the Model
With more than 100 engagements under our belts, we have learned that getting to the lab quickly is key to success. We encourage our customers to run three rounds of experiments in the first 3 months of any project. The model is a key, critical component in running effective workflows and helping design those rounds of experiments to test novel materials. It’s an indispensable tool in the modern materials scientist’s arsenal. But it doesn’t have to be perfect to add value. It doesn’t have to be perfect to help you learn more from your data and to help you pick the right materials to make. To use AI-driven experimental iteration well, researchers need a mindset shift, from scientist to engineer. Prototype early and often rather than finessing model metrics.
Models Guide
They illuminate pathways through complex design spaces and highlight promising directions
Experiments Validate
They test hypotheses and generate the real data that matters—actual material performance
Materials Deliver
They solve customer problems and generate revenue when they meet specifications
The real value only comes when you run the experiments and make the materials. The model, by itself, is not a crowning achievement—it’s a means to an end.