Astrochemical Forecasting

with Machine Learning
  • Kelvin Lee,
  • Jacqueline Patterson,
  • Andrew Burkhardt,
  • Michael McCarthy,
  • Brett McGuire

ISMS 2021—Talk RC03

Molecules in space

From origins of life to star formation

Pushing the boundaries of molecular complexity

See talk RC06 by Brett McGuire for the most up to date numbers!

What's next for astrochemistry?
Laboratory assays
Chemical models
Spectra don't come with labels!
Models poorly reproduce aromatics!
 Need a method of systematically identifying key species!

Cheminformatics
graph LR; Databases-->Structure-Retrieval-->Screening-->Query

High-throughput detection of desirable molecules from millions of candidates

Combine with machine learning to recommend molecules for study

Machine representations of molecules

Hand-picked features

Unsupervised representation learning

mol2vec

  • Unsupervised language model adapted for molecule strings.
  • SMILES encoding transformed into k-neighborhood chemical environment
  • Generate a dictionary mapping between environments and vectors

Molecule vector representations

Distance and similarity measures in Euclidean space

$\cos(\theta) = \frac{\mathbf{A} \cdot \mathbf{B}}{\vert \mathbf{A} \vert \vert \mathbf{B} \vert}$

Molecule vector representations

Visualization of chemical inventory clustering and trends
Where did we get our molecules?
Source Number of entries
ZINC 3,862,980
PubChem A 2,444,441
PCBA 437,929
QM9 133,885
NASA PAH database 3,139
KIDA 578
TMC-1 inventory 88

Total of 3,316,454 unique molecules used for mol2vec training

455,461 chosen from unsupervised clustering based on proximity to TMC-1 inventory

Astrochemical forecasting

Data pipeline for astrochemical forecasting

Reproducing chemical inventories

As simple as linear regression to custom kernels in Gaussian Processes

Outperforms conventional chemical models

Model performance across chemical complexity

Predicting the abundance of potential molecules

1507 unique molecules identified from 100 nearest neighbors per molecule in TMC-1

17x the number of currently detected species
Recommended molecules compliment existing KIDA network

Molecules for laboratory and astronomical study

Out of 1507 molecules...

59% are unsaturated

12% are pure hydrocarbons

43% contain are cyanides

New aromatics up to three rings

Under review—contact me for the full list!

Where to from here?

Apply ML approach to other prototypical sources

Infer abundance of radio-invisible molecules

Guide laboratory/observation/modeling efforts

Multi-output models
NASA-NNX13AE59G NASA-80NSSC18K0396 SAO REU Program under AST-1852268 and DoD ASSURE

Acknowledgements

Thank you!

github.com/laserkelvin

@cmmmsubmm

Google scholar

Finding new molecules to study by intuition alone is difficult.

mol2vec brings a new way to study astrochemical inventories.

Use ML to screen molecules

Identify candidates for study and predict their column density