Many computational tools are powerful once the structure, variables and target are already known. Some of the most interesting scientific problems begin earlier than that: what should be modelled, which microscopic states matter, and which representation is appropriate may themselves be unknown.
Why 22MI exists
In molecular and materials research, the available evidence is often incomplete. A structure may be unknown or non-unique, several reaction pathways may compete, an experiment may reveal only part of a system, or the relevant mechanism may change with conditions.
22MI works in this gap. We reconstruct plausible microscopic explanations, test them against physical constraints and experimental evidence, and turn successful methods into reusable computational workflows and software.
From incomplete evidence to testable models
We call this approach virtual reconstruction: building a computationally testable picture of what may be happening underneath incomplete observations.
The level of detail depends on the question. One problem may require atomistic simulation, another may begin from composition or spectroscopic data, and another may be better described as a reaction network. The model is chosen for the evidence available and the decision it needs to support.
In practice, this can mean identifying hidden reaction steps between observed states, predicting molecular behaviour before a unique microscopic structure is available, or connecting measured spectroscopic signals to the molecular states that produce them.
Founders
Liu, Xiangyue
PhD in Physics · Co-Founder
Computational scientist developing machine-learning and data-driven methods for extracting reliable molecular predictions from limited or heterogeneous scientific data. Her work connects machine learning with spectroscopy, physical models and uncertainty analysis, and uses experimental and literature data to build predictive molecular models. She also develops scientific databases and web-based research platforms that connect curated data, calculations and reusable research interfaces.
Machine learningSpectroscopyUncertainty analysisScientific databases & platforms
Wang, Weiqi
PhD in Physics · Co-Founder
Computational scientist developing methods to discover how molecular systems evolve and react under realistic conditions. His work combines molecular simulation, reaction dynamics and automated reaction-network construction to identify hidden states, competing pathways and condition-dependent mechanisms directly from atomistic trajectories, including how reaction networks change with temperature, pressure and composition. He also develops scientific software and computational workflows that automate simulation, network construction, analysis and validation.
Molecular simulationReaction dynamicsAutomated reaction-network modellingScientific software & platforms
Together, the founders cover both sides of the problem: reconstructing microscopic mechanisms from simulation and extracting reliable predictions from experimental and scientific data. Scientific software and research infrastructure connect these capabilities into reusable systems.
From research to reusable systems
22MI is built around a progression from scientific problem-solving to reusable computational capability:
- Scientific models. Reconstruct mechanisms, states and predictions from simulation, data and experiment.
- Reusable computational workflows. Turn successful research methods into automated, validated and traceable procedures that can be applied beyond a single case.
- Scientific software and platforms. Connect mature methods, structured data and calculations through tools and interfaces that make them easier to inspect, reuse and extend.
Principles
- Physics-grounded. Predictions should respect physical constraints, and the assumptions and limits of a model should remain explicit.
- Uncertainty-aware. What is measured, what is inferred and what remains unknown should remain distinguishable rather than being hidden behind a single prediction.
- Experiment-connected. Models should connect wherever possible to measurable observables, realistic conditions and evidence that can validate, refine or reject them.
Where this is going
Our long-term goal is to turn specialised scientific reasoning into reusable computational systems that can support research across chemistry, materials and molecular technology. The aim is not to replace scientific judgement, but to make difficult reasoning more reproducible, testable and scalable.