Accelerating Scientific Workflows with LLM Agents

Large Language Model (LLM) agents are transforming the way we approach the Design, Make, Test, and Scale (DMTS) steps in chemistry research. These AI-driven collaborators can extract insights from literature, model reaction conditions, and analyze data with unprecedented speed.

Example Reaction and Corresponding Prompts

Acetylsalicylic acid is synthesized via the esterification of salicylic acid with acetic anhydride, catalyzed by sulfuric or phosphoric acid under reflux. The reaction follows a nucleophilic acyl substitution mechanism, yielding aspirin and acetic acid as a byproduct. Purification is typically achieved through recrystallization from water or ethanol to remove unreacted reagents and side products.

35 Prompts to Guide the Process (DMTA)

Some prompts are conceptual or under evaluation. Fine-tuning the overall system -the interaction of the agents – requires further work.

#	Prompt	Answer (Example)	Involved LLM Agent	Involved Tool
Discovery Phase
1	Supervisor, initiate aspirin discovery workflow	Discovery workflow initialized	Supervisor	Paramus Routing
2	Retrieve molecules similar to aspirin	[Salicylic acid, methyl salicylate, benzocaine]	Data	Molecular Registry
3	Generate SMILES for acetylsalicylic acid	CC(=O)Oc1ccccc1C(=O)O	Open Cheminformatics (not LLM!)	RDKit
4	Compute physicochemical properties for aspirin	MW=180.16, logP=1.19	Open Cheminformatics	RDKit
5	Check inventory for salicylic acid and acetic anhydride	Available: Salicylic Acid (100g), Acetic Anhydride (250 mL)	Data	Inventory System
6	Review known aspirin reactions.	Fischer esterification, acetylation (standard synthesis)	Literature	Google
7	Summarize known risks of aspirin synthesis	Hydrolysis risk, moisture sensitivity	Literature	PubChem, ACS
Data Analysis Phase
8	Design a fractional factorial DOE for aspirin synthesis	Generated DOE with factors: Temperature (60-80°C), Catalyst (H₂SO₄)	DataScience	scikit-learn
9	Calculate moles from 5g salicylic acid	5 g corresponds to 0.0362 mol	Calculator	Python Interpreter
10	Analyze experimental results to find optimal temperature	Optimal temperature: 70°C	DataScience	Random Forest (scikit-learn)
11	Evaluate impurity correlation with catalyst amount	Higher catalyst correlates to fewer impurities (r²=0.85)	DataScience	Statistical Model
Experimentation Phase
12	Compute mass of acetic anhydride needed for reaction	7.39 mL required	Open Cheminformatics, Calculator	RDKit, Python Interpreter
13	Check availability of acid catalyst in the lab	Catalyst (Sulfuric acid, 50 mL) available	Data	Inventory Query
14	Predict reaction time for 95% yield	Reaction time ~45 min at 75°C	Literature, Cheminformatics, DataScience	Simple ML Model (Regression)
(hypothetical) 15	Monitor reaction temperature stability	Temperature stable at 75±1°C throughout reaction	Data	(2026: Paramus Collector) → IoT Sensor Integration
16	Confirm reaction completion using TLC data	TLC indicates no salicylic acid presence; reaction complete	DataScience	Image Analysis (OpenCV)
Optimization Phase
17	Suggest adjustments to reduce reaction impurities	Decrease temperature to 68°C and reduce catalyst by 10%	DataScience, Cheminformatics, Literature	DOE/Statistical analysis
18	Calculate yield improvement from temperature adjustment	Estimated yield increase: 4.5%	Calculator	Python Interpreter
19	List historical experiments related to yield optimization	Experiment IDs [THGR-201, THGR-434, AMBR-1001]	Data (from ELN!)	ELN
20	Verify inventory for scale-up materials	Materials sufficient for 10-fold scale-up	Data	Inventory
21	Compute cost impact of scale-up conditions	Cost reduction: 15% per gram product	Literature, Code	Python Interpreter
Literature Review Phase
22	Retrieve latest pharmacokinetics studies on aspirin	3 recent studies found, summarized in file	Literature	PubMed API
23	Summarize current safety guidelines for aspirin use	FDA guidelines updated 2023, dosage max 325 mg/day	Literature	Google
24	Check regulatory compliance for aspirin synthesis method	Complies with current chemical manufacturing standards	Literature	Google
25	Obtain current patent landscape for aspirin synthesis	No conflicting patents found	Literature	Google patent
26	Search adverse events databases for aspirin	Common adverse events: gastrointestinal irritation	Literature	Google
Computational Validation Phase
27	Optimize aspirin molecular geometry	Optimized structure converged at B3LYP/6-31G*	Computational Chemistry	PSI4
28	Compute infrared spectrum of aspirin	Major IR peaks at 1750 cm⁻¹ and 1650 cm⁻¹	Computational Chemistry	PSI4
29	Predict aspirin solubility via QSAR modeling	Predicted solubility in water: 3 mg/mL	DataScience, Cheminformatics	sklearn, SVM and ensemble models
30	Calculate transition states of aspirin hydrolysis	Activation energy calculated: ΔG‡ = 23.4 kcal/mol	Computational Chemistry	PSI4, OPT_TYPE set to TS
31	Verify predicted IR spectra against literature values	IR peaks match literature within ±5 cm⁻¹	Literature	Pubchem
Deployment Phase
32	Store optimized reaction protocol	Protocol saved as THGR-ASP-2025-01	Data	ELN (is also in Copilots file system)
33	Update inventory based on experimental consumption	Inventory updated: salicylic acid reduced by 50g	Data	Inventory UPDATE
34	Generate compliance report for QA	Report generated and ready for QA review	Literature	as .rd file in local file system
35	Prepare summary of synthesis and computational data for internal report	Summary report ready; includes experimental yields, computational validation, and literature references	Supervisor	as .rd file in local file system

Although not every agent from the example is fully productive yet, each one contributes targeted expertise that drives efficiency. By combining literature analysis, computational chemistry, and data insights, researchers gain a powerful toolset to streamline a development of aspirin synthesis. The following curated list of prompts illustrates how LLM agents can pave the way for a more robust and scalable DMTS process.