Our AI & ML Modelling Pipeline
A structured approach from data curation to production-ready, explainable predictive models
Curate & Enrich Scientific Data
Collect, clean, and harmonise experimental datasets from literature, databases, and project partners. Enrich with computationally derived atomistic descriptors to capture structural and physicochemical features beyond what experimental data alone provides.
Address Data Challenges
Handle class imbalance through synthetic data generation techniques, manage dataset heterogeneity, and ensure broad chemical or material coverage. Apply rigorous quality control and meta-analysis to build robust, modelling-ready datasets.
Build & Optimise Models with AutoML
Apply automated machine learning to systematically evaluate algorithms, optimise hyperparameters, and select the best-performing models. Use ensemble and consensus strategies from multiple modelling teams to improve accuracy and reduce bias.
Validate for Regulatory Acceptance
Rigorously validate models according to OECD principles for QSAR development, including defined endpoints, unambiguous algorithms, defined applicability domains, and appropriate measures of goodness-of-fit and predictivity.
Explain, Deploy & Share
Apply SHAP, permutation importance, and other explainability techniques to reveal key prediction drivers. Deploy validated models as user-friendly web services on the Enalos Cloud Platform and make datasets available through FAIR repositories.
Core AI & ML Capabilities
Technologies and methods that power our predictive modelling and decision-support services
Automated Machine Learning
Systematically evaluate and optimise ML algorithms to build high-performance models with minimal manual intervention, from feature selection through to hyperparameter tuning.
- Multi-algorithm benchmarking
- Automated feature selection
- Hyperparameter optimisation
- Ensemble & consensus strategies
Atomistic Descriptor Enrichment
Augment experimental datasets with computationally derived descriptors from crystal structures and molecular simulations to capture features invisible to standard characterisation.
- NanoConstruct & ASCOT toolboxes
- Surface & bulk atom descriptors
- Molecular & periodic table features
- Force field-based energetics
Explainable AI (XAI)
Make model decisions transparent and interpretable for experimentalists, regulators, and stakeholders using state-of-the-art explainability and interpretability techniques.
- SHAP & Shapley additive values
- Permutation importance analysis
- Feature contribution mapping
- Mechanistic insight extraction
See It in Action
Six real-world publications demonstrating how NovaMechanics applies AI and machine learning to nanosafety, drug discovery, and environmental science.
Explainable Prediction of Iron Carbide NP Cytotoxicity via Enalos Cloud
Developed data-driven workflows for iron carbide nanoparticle cytotoxicity risk assessment using atomistic descriptors from ASCOT, AutoML, and SHAP explainability — deployed as a free web service on the Enalos CHIASMA Cloud Platform.
Read the case studyIn Silico NP Toxicity Assessment Powered by Enalos Cloud
Built an automated ML workflow integrating atomistic descriptors, synthetic data generation, and a novel applicability domain method for nanosafety evaluation of Ag, TiO₂, and CuO nanoparticles — made available as SafeNanoscope web service.
Explore the case studyPredicting PPARγ Potency: Consensus Model & Deep Learning
Combined neural network binding affinity prediction with consensus ML classification for PPARγ antagonist activity — OECD-validated and applied to screen PFAS compounds for endocrine disruption potential via Enalos Cloud Platform.
View the case studyAutoML Prediction of Plant Length Responses to Nanoparticles
Applied rigorous data curation, atomistic enrichment, and automated ML to predict how nanoparticle exposure affects plant growth — achieving 85% accuracy with the XGBoost model deployed as CeresAI-nano web application.
Read the case studyPredicting Metal Oxide NP Cytotoxicity Using Isalos Analytics
Built a QNAR model for 24 metal oxide nanoparticles enriched with 62 atomistic computational descriptors, identifying 7 statistically significant drivers of cytotoxicity — deployed as a publicly available web service.
Explore the case studyRound Robin Consensus Prediction of Nanomaterials Zeta Potential
Four international research groups independently built ML models for zeta potential prediction of metal and metal oxide nanomaterials, then integrated them into a consensus scheme that outperformed any individual model.
View the case study