CHAOS -- A Consistent Large-scale Database for Sigma-Profiles and Other Molecular Descriptors

This paper introduces CHAOS, a large-scale, internally consistent open-access database containing high-accuracy sigma-profiles and diverse quantum-chemical descriptors for over 53,000 molecules, generated via a standardized wB97X-D/def2-TZVP workflow to support advanced molecular modeling and data-driven design across chemistry and materials science.

The Gist

Imagine you are a master chef trying to invent the perfect new recipe. To do this, you need to know exactly how every ingredient behaves: how it smells, how it reacts to heat, and how it mixes with other ingredients. In the world of chemistry, "ingredients" are molecules, and predicting how they mix is crucial for making everything from new medicines to better fuels.

For decades, scientists have used a special "fingerprint" called a $\sigma$-profile to understand these molecules. Think of a $\sigma$-profile as a molecular weather map. It doesn't just show you the shape of the molecule; it shows you where the "sunny" (positive charge) and "stormy" (negative charge) spots are on its surface. This map tells you if a molecule will stick to water, repel oil, or bond with hydrogen.

The Problem: A Messy Library

Until now, the problem was that scientists were trying to build a recipe book using weather maps from different sources.

• One scientist used a "sunny-day" calculator (Method A).
• Another used a "cloudy-day" calculator (Method B).
• A third used a "stormy-night" calculator (Method C).

Even though they were all measuring the same molecule, their maps looked different because the tools were different. Trying to combine these maps into one big, consistent library was like trying to build a house using bricks from three different factories that don't quite fit together. The result was a library that was too small and too messy to be truly useful for modern computer programs.

The Solution: CHAOS

Enter CHAOS (Computed High-Accuracy Observables and Sigma Profiles). The researchers at the University of Kaiserslautern have built a massive, perfectly organized digital library containing the weather maps for 53,091 different molecules.

Here is what makes CHAOS special, explained simply:

1. One Rule for Everyone: Instead of using different calculators, they used one single, high-precision method (a specific type of quantum physics calculation) for every single molecule. It's like if every brick in your house was made by the same factory, using the exact same mold. This means every map in the library is perfectly compatible with every other map.
2. More Than Just Maps: CHAOS doesn't just give you the weather map. It also provides a full "ID card" for each molecule, including:
   • The Shape: Exactly how the atoms are arranged.
   • The Sound: How the molecule vibrates (like its unique musical note).
   • The Magnetism: How it reacts to magnetic fields (like a compass).
   • The Solubility: How it behaves when dissolved in liquid.
3. The Scale: Previous libraries had about 2,000 molecules. CHAOS has 53,000. They expanded the chemical universe available to scientists by more than 10 times.

How They Did It (The Assembly Line)

The team didn't just guess the shapes of these molecules; they built them step-by-step on a supercomputer assembly line:

• Step 1: The Rough Draft. They started with a basic sketch of the molecule and used a fast, simple tool to find a decent shape.
• Step 2: The Polish. They used a smarter tool to check if the molecule could twist or bend into a better shape, discarding the awkward ones.
• Step 3: The Masterpiece. Finally, they used a super-precise, high-energy calculation to get the perfect shape and generate all the data.
• Step 4: The Catalog. They saved all this data into a neat digital file (JSON) that any computer can read.

Why This Matters

This database is a goldmine for Artificial Intelligence (AI).
In the past, AI trying to predict how chemicals mix was like a student trying to learn math with a textbook that had missing pages and conflicting answers. Now, with CHAOS, the AI has a complete, consistent, and massive textbook.

This allows scientists to:

• Design new solvents for greener industrial processes.
• Discover new medicines faster by predicting how drug molecules will interact with the body.
• Create better batteries and materials by understanding exactly how molecules behave together.

The Bottom Line

The CHAOS database is like giving the entire scientific community a universal translator and a massive, consistent encyclopedia of molecular behavior. It removes the confusion of "different calculators" and provides a single, reliable foundation for the next generation of chemical discovery. And the best part? It's free for anyone to use, download, and learn from.

Technical Summary

1. Problem Statement

The paper addresses critical limitations in the availability and consistency of σ-profiles (surface charge density distributions derived from quantum-chemical calculations), which are essential descriptors for:

• Thermodynamic Modeling: Used in models like COSMO-RS, COSMO-SAC, and hybrid group-contribution methods to predict activity coefficients and phase equilibria.
• Machine Learning (ML): Serving as physics-based molecular descriptors to predict thermophysical properties.

Key Challenges Identified:

• Scarcity: Existing public repositories (e.g., LVPP, VT-2005, NIST COSMO-SAC) contain only 1,400–2,300 molecules, limiting the training of robust ML models.
• Inconsistency: Existing datasets are generated using disparate quantum-chemical methods (different functionals, basis sets, and software). This introduces systematic shifts in σ-profile histograms, making data from different sources incomparable without complex reparameterization.
• Fragmentation: There is a lack of large-scale datasets that simultaneously provide σ-profiles alongside other high-quality quantum-chemical observables (e.g., NMR, IR, ideal-gas properties) generated under a single, unified protocol.

2. Methodology

The authors developed CHAOS (Computed High-Accuracy Observables and Sigma Profiles), a database containing 53,091 molecules. The data generation followed a rigorous, standardized multi-step workflow:

A. Data Source and Preprocessing

• Source: Molecules were selected from the Dortmund Data Bank (DDB), ensuring the availability of experimental thermophysical data for validation.
• Canonicalization: Initial MOL files were converted to canonical representations to remove atom-ordering inconsistencies.
• Conformer Generation:
  1. 3D Embedding: Used RDKit with the ETKDG algorithm.
  2. Force Field Optimization: Initial geometry optimization using the Universal Force Field (UFF).
  3. Semi-Empirical Refinement: An exhaustive conformer search using CREST with the GFN2-xTB Hamiltonian to capture non-covalent interactions (hydrogen bonding, dispersion) and identify low-energy minima within a narrow energy window.

B. High-Level Quantum-Chemical Calculations

All final calculations were performed using Gaussian 16 with a single, uniform protocol to ensure internal consistency:

• Theory Level: ωB97X-D (a range-separated hybrid functional with dispersion corrections) combined with the def2-TZVP (triple-ζ) basis set.
• Workflow Steps:
  1. Geometry Optimization: To find the global minimum.
  2. Harmonic Frequency Analysis: To confirm true minima (no imaginary modes) and derive vibrational properties.
  3. GIAO NMR: Calculation of Nuclear Magnetic Resonance shielding tensors.
  4. C-PCM Single-Point: Conductor-like Polarizable Continuum Model calculation to generate the tessellated COSMO surface, segment charges, and solvation energies.

C. Data Processing

• Raw C-PCM outputs were processed to generate σ-profiles (total and partial) following the open-source COSMO-SAC-dsp protocol.
• Ideal-gas thermodynamic properties (heat capacities, entropies) were derived from vibrational frequencies using the rigid-rotor harmonic-oscillator (RRHO) approximation.
• All data were structured into JSON files containing six categories: General, Structural, Electronic, Vibrational, NMR, and Solvation.

3. Key Contributions

• Scale and Scope: CHAOS expands the publicly available σ-profile space by more than an order of magnitude (53,091 molecules vs. ~2,000 in previous largest sets).
• Internal Consistency: Every molecule is calculated using the exact same ωB97X-D/def2-TZVP workflow, eliminating systematic errors caused by mixing different computational methods.
• Multi-Descriptor Integration: Unlike previous databases that focus solely on energies or specific properties, CHAOS provides a unified set of:
  • σ-profiles (for solvation/thermodynamics).
  • Vibrational data (IR spectra, ZPE, heat capacities).
  • NMR data (shielding tensors, magnetic susceptibilities).
  • Electronic descriptors (dipole moments, HOMO-LUMO gaps, polarizabilities).
  • Solvation data (cavity volumes, surface areas, non-electrostatic energies).
• Chemical Diversity: The dataset covers molecules with masses up to 400 amu and dipole moments up to 15 D, including diverse heteroatoms (O, N, S, halogens) and functional groups.
• Open Access: The database is freely available on Zenodo under a CC-BY-4.0 license, accompanied by the source code for the workflow on GitHub.

4. Results and Database Characteristics

• Molecular Distribution:
  • Mass: 2 amu (Hydrogen) to 394 amu (Iodoform).
  • Dipole Moment: 0 D (hydrocarbons) to 15.1 D (α-2'-deoxythioguanosine).
  • Composition: ~7% hydrocarbons; Oxygen is the dominant heteroatom, followed by Nitrogen, Sulfur, and Halogens.
  • Diversity: Tanimoto similarity analysis of random pairs shows that the vast majority of molecules have low structural similarity, confirming the dataset's broad coverage of chemical space.
• Validation:
  • Zero-point energies (ZPE) and dipole moments were validated against the NIST CCCBDB reference database, showing high agreement.
  • Entries with persistent imaginary frequencies after reoptimization are transparently flagged for user filtering.
• Comparison:
  • Vs. QM9/QM7-X: While QM9 (134k molecules) and QM7-X (42M molecules) are larger, they focus on energies/forces at lower levels of theory (e.g., B3LYP/6-31G*) and lack explicit COSMO/solvation data. CHAOS offers higher methodological accuracy (triple-ζ) and unique solvation/spectroscopic descriptors.
  • Vs. CCCBDB: CHAOS offers a much larger chemical space and includes solvation descriptors, which CCCBDB lacks.

5. Significance and Impact

• Foundation for Data-Driven Science: CHAOS provides the first large-scale, consistent foundation for training machine learning models for thermodynamic property prediction, bridging the gap between physics-based models and data-driven approaches.
• Model Reparameterization: The consistency of the data allows for the robust reparameterization and validation of COSMO-based models (COSMO-RS, COSMO-SAC) without the noise introduced by mixed computational protocols.
• Versatility: The inclusion of NMR, IR, and ideal-gas properties alongside σ-profiles makes the database valuable for multiscale modeling, solvent design, and algorithmic screening in chemical engineering and materials science.
• Community Resource: By releasing raw C-PCM outputs, the authors enable the community to generate custom σ-profiles using alternative post-processing protocols, fostering further research and extension.

In summary, CHAOS represents a significant leap forward in computational chemistry infrastructure, transforming σ-profiles from a niche, fragmented resource into a comprehensive, high-fidelity dataset capable of driving the next generation of thermodynamic modeling and molecular design.