The patent landscape: four technical clusters, 50+ documents
AI-driven process parameter optimization for injection molding scrap rate reduction is an active and geographically distributed field, with more than 50 patent documents spanning China, Japan, South Korea, the United States, Europe, and Brazil — of which approximately 25 are active or pending patents directly relevant to injection molding process optimization. The dominant assignees with multiple filings include BASF SE (JP, CN, BR, KR), IMFLUX INC. (US, WO, MX), Husky Injection Molding Systems (CA, EP, WO), LS Mtron (JP), and ITRI/Industrial Technology Research Institute (TW, CN).
The prevailing technical approaches cluster around four areas: (1) simulation-based closed-loop optimization, (2) reinforcement learning and iterative model training, (3) real-time melt and cavity pressure feedback control, and (4) multi-objective optimization using metaheuristic algorithms such as particle swarm optimization (PSO) and genetic algorithms. Chinese industrial assignees — including Kaos IoT Technology (Haier), Zhejiang University, and Beijing Baidu — represent a growing cluster of data-driven and cloud-connected approaches, reflecting a broader smart factory deployment agenda distinct from the machine-level focus of Western filings.
The AI injection molding process optimization patent landscape spans more than 50 documents across six jurisdictions (China, Japan, South Korea, the United States, Europe, and Brazil), with approximately 25 active or pending patents directly targeting scrap rate reduction through process parameter control.
The broader trend observable across the dataset is a migration from single-parameter closed-loop control — exemplified by the early Pennwalt viscosity-compensation patent filed in 1974 — toward multi-parameter, multi-objective AI frameworks that simultaneously optimize quality, cycle time, and energy consumption while maintaining dimensional tolerances. According to WIPO, cross-jurisdictional filing activity in AI-assisted manufacturing control has accelerated markedly since 2020, consistent with the clustering of filings in this dataset between 2021 and 2025.
Simulation-coupled AI: closing the loop between digital twins and physical machines
The most architecturally sophisticated approach to injection molding scrap reduction involves coupling physics-based simulation with machine learning to generate optimized process parameters before any physical shot is made — eliminating setup scrap at its source rather than measuring and correcting it after the fact. BASF SE’s 2023 patent defines a multi-step closed-loop in which an external processing unit first simulates the injection molding process using a simulation model combined with material-specific and machine-specific parameters; an optimization algorithm then determines predicted process parameters.
An adaptive digital twin in injection molding is a physics-based simulation model that is continuously recalibrated using actual process data from physical shots. BASF SE’s patented architecture feeds real workpiece measurement data back into the simulation model itself — so the virtual representation of the process improves with every production cycle, narrowing the gap between predicted and actual part quality.
These predicted parameters are fed to the physical machine, and actual part attributes are compared against optimization objectives. If the produced workpiece deviates from targets, process parameters are iteratively adapted until the part meets pre-defined tolerances. Crucially, actual process data is then fed back to recalibrate the simulation model itself — creating a self-improving adaptive digital twin. The Chinese counterpart filing from BASF SE explicitly states that “process data can be used to optimize the modeling process using machine learning models” and that “cloud-based digital twins of materials and injection molding processes” enable batch-specific material information to be incorporated into the simulation.
BASF SE’s patented adaptive digital twin architecture for injection molding uses AI-optimized predicted parameters from simulation before any physical shot is made, then feeds actual workpiece data back to recalibrate the simulation model — minimizing both initial setup scrap and ongoing process drift from material batch variation.
A related simulation integration approach is described in INGLASS S.P.A.’s WO patent, where CAE simulation results are electronically processed to generate initial machine parameters, and actual machine execution results are stored in a database alongside simulation-derived parameters. This database allows subsequent CAE simulations to be corrected as a function of the divergence between simulated and real outcomes — directly reducing the gap between virtual and physical process behavior that generates scrap during first-article qualification.
Space Solution Co.’s 2025 Korean filing extends the simulation approach further using an ANN-based (Artificial Neural Network) pressure prediction model trained on CAE injection analysis outputs. The system extracts pressure distribution pattern vectors from elements and nodes of the analysis object, then uses learned functional relationships to predict cavity pressure distributions under new conditions — enabling defect-prone pressure profiles to be identified and corrected before physical production begins. This approach aligns with the broader industry direction described by ISO in its smart manufacturing framework standards, where virtual process validation is positioned as a prerequisite for zero-defect production targets.
“Cloud-based digital twins of materials and injection molding processes enable batch-specific material information to be incorporated into the simulation — preventing the scrap spikes that characterize manual process restarts after a resin change.”
Nissei Plastic Industrial’s 2024 filing addresses a specific and frequently overlooked scrap trigger: resin batch changes. The system automatically recalculates molding conditions to match the melt characteristics of a new material batch to those of the original baseline — a targeted intervention for the documented problem of batch-switch scrap spikes that generic AI optimization systems are not designed to handle.
Explore the full patent landscape for AI injection molding optimization in PatSnap Eureka — search, filter, and compare 50+ filings by assignee, jurisdiction, and claim type.
Explore Patent Data in PatSnap Eureka →Real-time melt and cavity pressure control: the most mature scrap-reduction mechanism
Real-time melt pressure and cavity pressure feedback is the most mature and commercially deployed technical cluster for injection molding scrap rate reduction, with IMFLUX INC. — a Procter & Gamble-affiliated company — holding the dominant patent position across US, WO, and MX jurisdictions. The core mechanism establishes an optimal melt pressure curve from a baseline cycle and then adjusts injection pressure in subsequent cycles to force the monitored pressure to follow this reference trajectory.
IMFLUX INC.’s patented melt pressure control system for injection molding uses machine learning algorithms trained on data from multiple machines, molds, and melt materials to determine modifications to an optimal pressure curve — preventing out-of-tolerance shots proactively during process disturbances rather than detecting defective parts after ejection.
IMFLUX’s filing explicitly describes the integration of machine learning algorithms to determine modifications to the optimal melt pressure curve, with training data drawn from multiple machines, molds, and melt materials to correlate part quality outcomes with pressure profiles. By maintaining the pressure profile within the learned optimal window, out-of-tolerance shots — and therefore scrap — are prevented proactively rather than detected after the fact. This is a fundamental shift in quality control logic: from inspection-after-production to prevention-during-production.
IMFLUX’s complementary autotuning architecture takes a model-based approach to determining initial control parameter values from a machine/mold/material model, then executes a run of injection cycles while measuring operational parameters. When any operational parameter exceeds a defined threshold, control parameters — including the melt pressure profile and PID gain values — are automatically adjusted for subsequent cycles. This threshold-triggered tuning prevents extended runs of defective parts while the process drifts, directly limiting cumulative scrap volumes.
LS Mtron’s AI-based injection molding system addresses the disturbance rejection problem specifically: when a defective product is produced due to process disturbance, the system acquires current injection state data including viscosity profile and injection pressure values, inputs these into a trained “molding quality maintenance model,” and determines whether quality is being maintained. If not, the molding condition setting unit automatically generates corrected conditions to restore quality — eliminating the sustained defect runs that characterize manual intervention-dependent production lines. Research published by IEEE on industrial closed-loop control systems confirms that automated disturbance rejection reduces mean defect run length by orders of magnitude compared to operator-triggered correction.
Across all real-time pressure control patents in this dataset, the shared quality logic is prevention rather than detection: the AI system maintains the process within the known-good pressure window cycle-by-cycle, so out-of-tolerance parts are never produced — rather than being identified and rejected after ejection. This eliminates both the scrap material cost and the inspection cost associated with post-ejection quality checks.
ML model architectures for parameter recommendation and multi-objective optimization
A distinct group of patents focuses on the ML model architectures and optimization algorithms that translate historical process data into actionable parameter recommendations — addressing the problem of parameter setup from scratch for new molds or materials, where the absence of historical data is the primary source of first-article scrap. ITRI’s model-based machine learning system couples an injection molding process simulator with a reinforcement-learning-style optimization loop: a molding condition optimizer proposes parameter sets, the simulator evaluates whether they produce acceptable outcomes, and parameters are iteratively refined over training rounds.
The ITRI system explicitly tracks that, as training rounds increase, the number of parameter adjustments needed per round to achieve a simulated good part converges — meaning the model accumulates knowledge of the process space and navigates it more efficiently over time. This convergence behavior is the quantitative signal that the model has learned enough about the process to reliably generate near-optimal parameters without physical trial shots.
Steer Industrial (Shanghai)’s 2024 patent trains an AI model on historical injection data that explicitly includes both “yield parameters” (good product rate) and “adjustment records,” penalizing the model for parameter choices that historically required multiple correction cycles. The system also determines the optimal timing for parameter re-optimization based on real-time machine data — avoiding wasted optimization cycles between legitimate re-optimization triggers and ensuring the AI intervenes only when process drift makes re-optimization necessary.
PatSnap Eureka lets you search and analyse the full claims of every patent in this dataset — compare architectures, identify white space, and track assignee filing velocity.
Analyse Patents with PatSnap Eureka →Multi-objective PSO: dimensional compliance as a hard constraint
Kaos IoT Technology’s 2024 patent introduces a particularly practical multi-objective architecture for injection molding scrap rate reduction: three pre-trained linear regression models (production time model, energy consumption model, and dimensional error model) operate as sequential filters. Candidate parameter sets are first screened by the dimensional error model to reject configurations that would yield out-of-tolerance parts, and only passing candidates are then optimized for time and energy by particle swarm optimization (PSO). This architecture treats dimensional compliance as a hard constraint rather than a weighted objective — guaranteeing that no parameter combination reaching the machine would violate dimensional tolerances, regardless of throughput or energy pressures.
Kaos IoT Technology’s multi-objective injection molding optimization system uses a dimensional error model as a hard pre-screening filter before particle swarm optimization (PSO) — ensuring that no parameter set reaching the injection machine would violate dimensional tolerances, regardless of throughput or energy optimization pressures.
A further development from the same assignee incorporates a large language model (LLM) to parse natural-language user requests for parameter recommendations — a novel interface layer that lowers the barrier to AI-assisted process setup for non-expert operators. Beijing Baidu’s 2024 filing specifically trains a process parameter tuning model on injection pressure, injection volume, and injection speed per segment — parameters directly linked to short-shot and flash defects, the two most common dimensional scrap causes in high-volume injection molding. The integration of LLM interfaces with process optimization models is consistent with the broader trajectory described by NIST in its smart manufacturing reference architecture, where natural-language human-machine interfaces are identified as a key enabler of AI adoption in production environments.
Model-free optimization: scrap reduction without historical data
Zhejiang University’s 2024 US patent takes a fundamentally different approach that avoids surrogate model construction entirely. Rather than training on historical data, the method calculates gradient directions iteratively using an iterative gradient estimation method and applies an adaptive moment estimation (Adam) algorithm to allocate adaptive step sizes to each parameter. This enables convergence to near-optimal parameters through online iteration without requiring extensive offline training data or a pre-built model of the process — significantly reducing the number of physical test shots required during new mold qualification, where the absence of historical data is the primary barrier to rapid, scrap-free process setup.
Husky Injection Molding Systems’ 2025 WO patent takes a system-level approach: the controller selects from a library of optimization models based on the operational parameter to be improved, enabling targeted optimization of whichever parameter is currently driving scrap. This model-library architecture connects AI capability to specific production quality goals in a way that is accessible to operators without deep ML expertise — a practical deployment pattern for large-scale injection molding operations managing multiple molds and materials simultaneously. Husky’s approach reflects the innovation intelligence principle that AI tools must be deployable by domain experts, not just data scientists, to achieve production-scale impact.
Key players, emerging trends, and what comes next
The innovation landscape for AI-driven injection molding scrap reduction is consolidating around five identifiable strategic positions, each reflecting a different theory of where the most tractable scrap reduction opportunity lies. Understanding these positions is essential for R&D teams deciding where to invest in process optimization capability — and for IP professionals assessing freedom-to-operate and white-space opportunities.
BASF SE occupies the materials-intelligence position: their closed-loop simulation architecture integrates cloud-based digital twins with physical machines, with filings across JP, CN, BR, and KR jurisdictions. The cross-jurisdictional breadth signals a strategic intent to establish the digital twin architecture as a standard across global injection molding supply chains — particularly relevant for BASF’s polymer customers who operate injection machines in multiple geographies. The PatSnap Insights blog has previously documented how materials companies are increasingly filing process patents to protect application-layer value beyond the material itself.
IMFLUX INC. (Procter & Gamble) holds the dominant position in real-time pressure-profile-based adaptive control, with multiple US, WO, and MX filings covering melt pressure tracking, cavity pressure control, PID autotuning, and compression rate optimization. Their systematic approach to characterizing the optimal compression rate window for a given material — and then finding machine operating parameters that stay within that window — represents the most mature and commercially validated technical approach in the dataset.
LS Mtron (Korea) focuses on the disturbance rejection problem, with multiple JP-jurisdiction filings covering both rapid initial condition generation from mold specifications and real-time disturbance-triggered condition correction. Their approach is particularly relevant for high-mix injection molding environments where frequent mold and material changes are the primary source of scrap spikes.
The Chinese industrial IoT cluster — Kaos IoT/Haier and Beijing Baidu — represents a distinct deployment philosophy: multi-model pipelines designed for heterogeneous fleets of injection machines in smart factory environments, with LLM-based parameter recommendation interfaces that make AI-assisted process setup accessible to non-expert operators. This cluster is growing rapidly and is likely to become the dominant filing source within the next two to three years as Chinese smart factory deployments scale.
“The broader trend is a migration from single-parameter closed-loop control — exemplified by the 1974 Pennwalt viscosity-compensation patent — toward multi-parameter, multi-objective AI frameworks that simultaneously optimize quality, cycle time, and energy consumption while maintaining dimensional tolerances.”
ARBURG GmbH‘s knowledge-based procedural control approach — connecting machine control to a specialized expert knowledge base and using part geometry and sprue geometry to calculate starting process parameters — represents a hybrid of rule-based and data-driven approaches that is particularly relevant to first-article qualification without scrap-generating trial runs. This approach is complementary to, rather than competing with, the ML-based architectures described above: rule-based systems provide the starting point, and ML-based systems refine from there.
The overall trajectory of the dataset is clear: scrap rate minimization is no longer a standalone objective in injection molding process optimization. It is embedded as one constraint within a broader multi-objective production optimization agenda that simultaneously targets cycle time, energy consumption, and dimensional compliance. The technical implication is that future AI systems for injection molding will need to handle all four objectives concurrently — and the patent filings from 2023 to 2025 suggest that several assignees are already building toward this capability.