Predict Chemical Reaction
Outcomes Before You React
Chemdem uses AI to predict reaction success and yield before you even start — powered by real squarate-amide data and IBM RXN.
It started with a real chemistry problem, not a trend.
Chemdem began from a challenge inside a university research group. When researchers wanted to make a molecule, they could not easily know in advance whether the reaction would work or what yield to expect. The goal became simple: turn trusted lab knowledge into a tool that helps researchers make better decisions before spending time, reagents, and lab hours.
Built between chemistry, code, and design.
Chemdem is being built from a rare position: pharmaceutical science, organic chemistry, software development, and UX design coming together. The aim is not only to make the system technically work, but to make it usable and understandable for chemists at the bench.
Chemistry-first
Designed around real molecules, real reactions, and real research constraints.
Software-built
Structured with modern tools such as Python, Node.js, RDKit, and clean web interfaces.
Made for trust
Every prediction should show its reasoning, confidence, and references.
Starting narrow on purpose.
Chemdem's first focus is squaric acid chemistry, specifically amines reacting with diethyl squarate to form squaric acid monoamides. The scope is narrow because trust comes before scale. Before Chemdem expands to more reactions, it must prove the approach on chemistry the research group understands deeply.
The real wall is not the algorithm. It is the data.
The hardest part of Chemdem is turning chemistry from papers into reliable machine-readable data. Molecules need to be converted into SMILES, checked, verified, and labelled correctly. A single wrong structure can poison the prediction system, so data quality is treated as the foundation of the whole project.
Hover any step to see what happens there.
Chemdem is not pretending to be magic.
Chemdem is currently a transparent algorithm, not a fully trained machine-learning model. With a small verified dataset, a neural network would not be the honest tool. Instead, Chemdem reasons by analogy: it compares a new molecule to known reactions and uses the closest real examples to estimate outcome and yield.
“Chemdem does not learn chemistry yet. It remembers it.”
A prediction system that shows its working.
Honesty is the core feature.
A prediction number alone is not enough. Chemdem is designed to show where the answer came from, how confident it is, and when the system is unsure. Chemists can also flag results for human review, helping the dataset improve over time.
The future of Chemdem is built in phases.
Curated dataset + transparent algorithm
Chemdem is currently focused on verified data, rule-based logic, and explainable prediction.
Machine learning when the data is ready
Once the dataset grows into hundreds of verified reactions, Chemdem can move toward a real trained model that learns from trusted chemistry data.
Open research contribution network
The long-term goal is a human-curated, referenceable platform where research groups can contribute verified data and improve future predictions.
Open-source, but earned step by step.
Chemdem is open-source in spirit and currently scoped around university research data. The project should not claim universal chemistry prediction too early. It should earn scale through verified contributions, better references, and stronger validation.
Chemdem is still being built.
The future depends on better data, stronger validation, and more researchers challenging the system until it earns trust.