AI Agentic Workflow for Virtual Screening of Novel Semiconductor Dopants

This workflow automates the high-stakes, computationally intensive search for novel semiconductor dopants. It replaces a manual, sequential process of literature review, simulation setup, job submission, and result analysis with an orchestrated, agentic system. The operational upside is a 10-20x acceleration in candidate screening, directly reducing lab iteration costs and shortening time-to-market for advanced power electronics, RF devices, and next-gen nodes. Savings come from eliminating weeks of engineer toil per candidate and optimizing expensive HPC resource utilization.

Implementation integrates with HPC job schedulers like Slurm, simulation software APIs (VASP, Quantum Espresso), and materials databases. The orchestrator, built on LangGraph, manages specialized agents for structure generation, convergence monitoring, and property ranking. Critical controls include automated validation of simulation convergence, exception routing for failed jobs, and a mandatory human-in-the-loop approval gate before final candidate promotion. Observability is built into each agent step, logging computational parameters, energy outputs, and ranking rationale for full auditability and model refinement.

The workflow automates the repetitive, expert-intensive process of setting up, running, and analyzing thousands of Density Functional Theory (DFT) simulations to evaluate dopant incorporation and activation energy. Operational savings come from eliminating manual job submission, file manipulation, and data extraction, allowing a single engineer to manage a candidate pool orders of magnitude larger. The core architecture is an agentic orchestration layer, typically built with LangGraph or CrewAI, that sequences tasks and manages state across fragmented HPC and data systems.

Implementation integrates with HPC job schedulers (Slurm, Kubernetes), simulation software APIs (VASP, Quantum ESPRESSO), and lab systems (ELN, LIMS). Controls are critical: each agent's outputs undergo validation checks against physical plausibility bounds before progression. A human-in-the-loop gate reviews top-ranked candidates and ambiguous results before lab commitment. Observability is provided via a centralized dashboard tracking job status, agent decisions, and simulation convergence, ensuring full auditability and enabling rapid diagnosis of pipeline failures.

Phase 1 establishes the core data pipeline and a single-agent screening loop. We integrate with your existing VASP or Quantum ESPRESSO HPC schedulers and Materials Project APIs to run DFT calculations for a curated list of 50-100 candidate dopants. A ranking agent evaluates formation energy and electronic band structure impact, populating a centralized results database. This initial phase delivers a working prototype in 8-10 weeks, validating the architecture and providing immediate ROI by automating the manual setup and collation of simulation jobs.

Phase 2 introduces multi-agent orchestration using LangGraph to manage parallel screening streams. Specialized agents handle candidate generation from generative models, simulation job monitoring, and post-processing analysis. A critical human-in-the-loop review gate is implemented before high-confidence candidates are promoted for experimental validation. This phase scales throughput to thousands of candidates and integrates the workflow into your team's existing review and decision processes, governed by configurable confidence thresholds and audit logs.

Scaling a virtual screening workflow from prototype to production requires a governance layer that preserves scientific control while automating repetitive tasks. The operational upside comes from compressing candidate evaluation from weeks to hours, but the risk lies in misclassifying a promising dopant or missing a critical failure mode. Implementation must integrate with HPC schedulers (Slurm, Kubernetes), DFT software (VASP, Quantum ESPRESSO), and material databases, while routing low-confidence predictions and novel edge cases to a human-in-the-loop review queue before final ranking.

The rollout is sequenced, starting with a shadow mode where agent predictions are logged but not acted upon, building trust in the automated ranking logic. Controls include predefined confidence thresholds, automated validation of simulation convergence, and immutable audit logs linking each candidate to its simulation parameters and review status. Monitoring via Prometheus/Grafana tracks job success rates, scientist review latency, and prediction drift, ensuring the workflow delivers throughput gains without compromising the integrity of the material discovery pipeline.

The workflow's value hinges on ingesting high-fidelity material data from sources like the Materials Project, proprietary DFT software (VASP, Quantum ESPRESSO), and lab validation systems. Orchestrators like LangGraph manage agents for simulation job submission, convergence monitoring, and property extraction. This architecture automates the manual chaining of simulations and data wrangling, compressing screening cycles from months to days and focusing experimental budgets on the most promising dopant candidates identified by ranking logic.

Integration with lab information management systems (LIMS) like Benchling is critical for routing high-potential candidates to physical validation. The workflow must embed controls for data quality checks, exception handling for failed simulations, and audit trails for regulatory or IP documentation. Implementation requires APIs to HPC schedulers (Slurm, Kubernetes), simulation software, and the results data lake, ensuring the system operates as a closed-loop, decision-support engine rather than a disconnected set of scripts.