Publications, communications and archives.

Advancements and challenges in large-scale atomic force microscopy image stitching for semiconductor devices

Combine high-speed AFM scans into large-area maps and analyze them in Smart3, in collaboration with Infinitesima.

Metrospection: holistic AI-driven process control software platform dedicated to metrology and defectivity for patterning process at die and wafer levels

Combine measurements and defect detection, from patterns to wafers, in an open platform configured by the engineer.

Open collaborative AI platform to support process optimization for semiconductor industry with digital twin

Connect data cleaning, modeling and process optimization while retaining the ability to integrate proprietary algorithms.

Augmented Engineer: A No-Code AI Application Layer for Bridging the Gap Between Plasma Simulation and Real-World Processing

A no-code AI layer to help engineers connect simulation with process data and explore optimization trade-offs, with STMicroelectronics.

No DOI supplied for this reference.

Scalable and adaptable structural metrology for AR/VR device images based on deep learning

Measure slanted gratings in TEM and structures in CD-SEM with workflows that evolve alongside Meta’s process development.

Application of deep-learning based techniques for automatic metrology on scanning and transmission electron microscopy images

The foundations of Pollen’s AI metrology: annotation, detection, segmentation and measurements from user-provided examples.

Automatic classification of C-SAM voids for root cause identification of bonding yield degradation

Detects, classifies and measures voids automatically in full 300 mm acoustic-microscopy (C-SAM) wafer images of directly bonded wafers for CMOS image sensors. Defect classes are tied to their process origin to find the root cause of bonding yield loss.

Process optimization assistant powered with AI driven digital twin: multiobjective processes use case

Adds multi-objective optimization to the AI digital twin. Instead of one weighted figure of merit, it returns a set of optimal recipes (Pareto front) and lets the engineer choose the trade-off. Demonstrated on a triple-junction solar-cell efficiency case.

Digital twin technology introduction to increase yield of processes

Introduces Pollen's data-driven digital twin. AI models trained on experimental data, optionally enriched by physical simulation, simulate new recipes and optimize the process under constraints set by the engineer. The paper also studies how sampling strategy and training-set size affect accuracy.

Application of machine learning-based metrology for writer main pole CD measurement by CDSEM

Deep-learning metrology on CD-SEM images of hard-disk writer heads across four designs. Models trained across designs handle large process variations and defective images that traditional methods cannot. Development and execution cycles are much shorter.

Introducing machine learning-based application for writer main pole CD metrology by dual beam FIB/SEM

Compares machine-learning metrology on in-line FIB/SEM cross-sections with Western Digital's legacy machine vision, on recipe-writing time and measurement uncertainty. Enriched training sets make the models robust from new-product introduction to high-volume manufacturing, and one model can measure several designs.

A generic deep-learning-based defect segmentation model for electron micrographs for automatic defect inspection

A single generic model, trained on defects annotated in SmartDef3 from several industrial applications, detects, segments and measures defects in SEM/TEM images. It also works on new use cases without retraining.

Deep-learning-driven automation of inner and outer object segmentation in SEM/TEM imaging for semiconductor metrology

A fine-tuned transformer model segments multi-material objects (inner and outer parts) starting from a single bounding box. It reaches 89% recall and 80% precision for detection, and R² of 0.82–0.92 on inner and outer dimensions, well above previous methods.

Deep learning aided tool for fast and accurate segmentation of multi-part semiconductor features

An annotation tool where the user draws a box and the AI segments the object and the material regions inside it. Models trained on its annotations perform as well as models trained on manual annotations, while annotation is about 30× faster with less user bias.

Deep learning contour-based method for semi-automatic annotation of manufactured objects in electron microscopy images

A contour-based deep model with a loss function built on electron-microscopy physics annotates objects from a bounding box. Measurements from its contours reach under 10% mean relative error and over 90% correlation with manual annotation.

Edge detection using deep learning pipelines for TEM and SEM metrology by proposing generic and specific approaches

Deep-learning edge maps for low-contrast high-resolution TEM and SEM images. Users can take a generic ready-to-use model or tune one on their own data. Both are benchmarked on gates and wires for accuracy and time-to-recipe.

Next generation of decision making software for nanopatterns characterization: application to semiconductor industry

The early vision of Pollen's platform: a software infrastructure that plugs in specific business modules, customizes algorithms per R&D team, and handles CD, roughness, classification and defects in one step.

Using machine learning technology to accelerate the development of plasma etching processes

A machine-learning framework that works alongside experts. It learns from the engineer's measurement recipes on TEM, cross-section SEM and SEM images to automate measurements during etch process development.

Accelerate the analysis and optimization of lamellar BCP process using machine learning

Automated CD-SEM analysis of block-copolymer self-assembly to screen chemical formulations. Machine-learning models then predict process outcomes and estimate the process window under expert supervision.

Machine Learning assistant technology to facilitate Fin and 3D memory measurements on SEM and TEM images

A machine-learning measurement pipeline driven by the engineer's "measurement recipe" for Fin and 3D-memory structures. It includes physical modelling of the SEM signal to locate blurred edges.

Cognitive learning: a machine learning approach for automatic process characterization from design

Uses the device design as input to train detection and measurement quickly. It is aimed at both machine-learning experts and non-experts and connects to process optimization.

Optimization of Complex Process, Based on Design Of Experiments, a Generic Methodology

A generic AI process-optimization workflow covering data cleaning, dimension reduction, modelling and multi-objective optimization. It was validated with Aledia's process experts on microLED manufacturing and is transferable to other materials processes.

No DOI supplied for this reference.

Collaborative and Holistic AI driven process control software platform to accelerate IC Manufacturing

Vision of a collaborative AI platform for heterogeneous metrology and inspection data (SEM, TEM, optical, acoustic, AFM). Objects are learned from a few examples, measurements are defined with geometric constraints, processing stays local for confidentiality, and an SDK supports customer customization.

No DOI supplied for this reference.

Introduction of a customizable AI cluster to accelerate accurate Metrology and Defectivity for Yield Enhancement and to build advance process Digital Twin

An AI cluster of metrology, defectivity and process-optimization pipelines. It has an API for automation and connection to the manufacturing execution system, covers R&D to production, and works with any microscope type (SEM, TEM, C-SAM, CD-SEM).

No DOI supplied for this reference.

Fast and Versatile Segmentation of Semiconductor materials based on a single example for metrology

Segmentation that needs only one annotated object. It reuses previously learned knowledge at prediction time, so non-expert users get results without tuning any model parameters.

No DOI supplied for this reference.

Deep learning performances over classical approaches in metrology domain applied on electronic microscopy images

Benchmarks Pollen's first-generation tool (machine learning plus image processing) against a deep-learning-only tool on wires, pillars and gates. Deep learning generalizes better and reaches high metrology performance with 6 or fewer training images (about 50 objects).

No DOI supplied for this reference.