Cold start
A cold start refers to the startup of an internal combustion engine after prolonged inactivity, when its components, including oil and coolant, are at ambient temperature well below the engine's typical operating range of around 80–100°C.[1][2] This occurs routinely in vehicles but is particularly demanding in subfreezing conditions, where thickened lubricants delay oil circulation, fuel struggles to vaporize effectively, and battery output diminishes, collectively straining the starting mechanism and initial combustion.[3][4] The phenomenon drives disproportionate engine wear and pollutant output relative to warm operation, as boundary lubrication fails briefly until oil pressure builds and surfaces align under heat expansion, with friction peaking on components like pistons, bearings, and cams.[3][5] Cold starts also elevate unburnt hydrocarbons, carbon monoxide, and nitrogen oxides until the three-way catalyst activates, often comprising 50–80% of trip-total hydrocarbons and significant CO fractions in urban driving cycles despite lasting under a minute.[6][7] Mitigation strategies have evolved from manual chokes and carburetor enricheners to electronic fuel injection for precise air-fuel ratios, glow plugs in diesels, and auxiliary preheaters like block or oil pan heaters that precondition fluids without idling.[4] Synthetic low-viscosity oils further reduce startup drag, while regulatory tests like FTP-75 simulate cold starts to enforce emission limits, underscoring their role in air quality debates.[3][8] Though modern engines tolerate cold starts robustly, excessive frequency accelerates degradation, prompting recommendations to minimize them via garaging or preconditioning in harsh climates.[5]Automotive engineering
Definition and mechanisms
A cold start in automotive engineering denotes the startup of an internal combustion engine after prolonged inactivity, where components such as coolant, oil, and combustion chamber surfaces have cooled to ambient temperature without prior thermal preconditioning.[2] This phase features markedly lower thermal efficiency than steady-state operation, attributable to heightened frictional and pumping losses from elevated lubricant viscosity, incomplete combustion, and accelerated heat dissipation to chilled metal surfaces.[9] Fuel consumption rises by as much as 7% during cold starts due to these inefficiencies.[9] The fundamental mechanism commences with the battery supplying electrical power to the starter motor, which rotates the crankshaft at cranking speeds of roughly 150-250 RPM to compress the air-fuel mixture sufficiently for ignition.[10] In both gasoline and diesel engines, cold conditions exacerbate oil viscosity, elevating friction mean effective pressure (FMEP) to approximately 10 bar initially at -20°C, which declines to 7 bar after about 100 crankshaft revolutions as localized heating occurs.[10] Pumping losses intensify from denser intake air and restricted exhaust flow through cold catalysts, while blowby—leakage of charge past piston rings—accounts for up to 10% mass loss at low speeds (e.g., 200 RPM), cooling the mixture by roughly 100°C and impeding pressure buildup.[10] In gasoline engines, mechanisms center on vaporization deficits: cold intake walls promote fuel condensation and wall wetting, yielding incomplete mixing and misfires until the engine control unit (ECU) enriches the air-fuel ratio to sustain combustion, often prolonging the warm-up phase with elevated hydrocarbon emissions from unburned fuel.[9] Heat transfer from the flame to cylinder walls further diminishes indicated mean effective pressure (IMEP), which requires exceeding 500 RPM to surpass frictional thresholds for self-sustaining operation.[10] Diesel engines face amplified challenges owing to reliance on compression-induced autoignition, where reduced cranking speeds curtail peak pressures and temperatures, extending ignition delay and risking up to 28 misfires per cycle at -20°C, each emitting around 1500 mg of hydrocarbons.[10] Glow plugs mitigate this by heating chamber surfaces to 850°C or higher, optimizing fuel atomization and slashing misfires below 5% while shortening ignition delay to 8-10 ms; retarded injection timing (e.g., 3.5° before top dead center) and increased pilot fuel quantities (30-50 mg) further boost IMEP to 7 bar for reliable firing.[10] Time to initial combustion varies from 0.3 seconds at 0°C to 1.2 seconds at -20°C, with full idle stabilization extending to 36 seconds in severe cold, underscoring the interplay of volatility-limited vaporization and thermal inertia.[10]Challenges and effects
Cold starts in automotive engines present significant challenges primarily due to reduced fluid mobility, impaired combustion efficiency, and diminished electrical performance at low ambient temperatures. In gasoline engines, fuel evaporation is hindered by cold intake air and manifold surfaces, leading to incomplete vaporization and poor air-fuel mixing, which can result in misfires or extended cranking times.[4] Diesel engines face even greater difficulties, as the higher compression ignition threshold requires elevated temperatures for autoignition; below -10°C (14°F), cetane numbers drop, causing delayed combustion and visible white smoke from unburned hydrocarbons.[11] Battery capacity and cranking speed decline by up to 50% at 0°F (-18°C) compared to 80°F (27°C), exacerbating starter motor strain and increasing the risk of failure after repeated attempts.[12] Lubricating oil viscosity rises dramatically in cold conditions, thickening to impede flow and providing inadequate initial lubrication to pistons, bearings, and valvetrain components.[13] These challenges manifest in several adverse effects on engine operation and longevity. Emissions spike during the cold start phase, with hydrocarbons (HC) and carbon monoxide (CO) levels up to 10-20 times higher than warm operation due to inefficient catalytic converter performance, which requires 200-400°C (392-752°F) to activate fully.[14] Fuel consumption increases by 10-20% in the initial minutes post-start, as the engine control unit enriches the mixture to compensate for poor combustion, leading to higher overall trip CO2 output in short drives dominated by cold phases.[15] Engine wear accelerates from boundary lubrication conditions, where metal-to-metal contact occurs before oil pressure builds, potentially shortening component life by promoting abrasive particles and fatigue in cylinder walls and rings.[16] In diesel applications, prolonged cranking without ignition heightens starter solenoid overheating and glow plug degradation, while incomplete combustion deposits soot in injectors, compounding long-term efficiency losses.[13] Overall, cold starts contribute disproportionately to urban fleet emissions inventories, accounting for up to 70% of trip HC in modal testing cycles despite comprising only the first 1-2 km of travel.[17]Mitigation techniques
Engine block heaters, typically electric immersion devices installed in the engine's coolant jacket, preheat the engine coolant and block to temperatures around 30–50°C (86–122°F) prior to starting, reducing cranking time and friction-related wear by circulating warm fluid through the system.[18] These heaters can decrease cold start wear by up to 50% by maintaining oil fluidity and minimizing thermal stress on components like pistons and bearings.[19] They are particularly effective in sub-zero conditions, where unheated engines risk incomplete lubrication during initial revolutions, and are standard in vehicles operating in regions with average winter lows below -10°C (14°F).[20] For diesel engines, glow plugs serve as resistive heating elements in each cylinder, elevating combustion chamber temperatures to 500–900°C (932–1652°F) within seconds to vaporize fuel and ignite the air-fuel mixture despite low ambient temperatures.[21] Modern self-regulating glow plugs activate via engine control units based on coolant temperature, often requiring 2–15 seconds of preheating below 9°C (48°F), which shortens cranking duration and lowers hydrocarbon emissions by improving initial combustion efficiency.[22] Failure or inadequate glow plug function can extend starts by 5–10 seconds or more in cold weather, increasing starter motor strain.[23] Synthetic engine oils, formulated with uniform molecular structures, exhibit superior low-temperature viscosity (e.g., 0W grades pour at -35°C/-31°F or lower versus -30°C/-22°F for conventional 5W equivalents), enabling faster oil pump priming and bearing lubrication during the critical first 20–30 seconds post-start.[24] This reduces startup friction by 20–30% compared to mineral oils, as measured in pumpability tests, and supports quicker achievement of full oil pressure, thereby cutting wear on crankshafts and camshafts.[25] In diesel applications, synthetic formulations also resist fuel dilution from incomplete cold combustion, maintaining additive efficacy against gelling.[26] Battery heaters, such as silicone pad or wrap-style units drawing 40–75 watts, maintain electrolyte temperatures above 0°C (32°F) to preserve cold cranking amps (CCA), which can drop 50% or more at -18°C (0°F) in lead-acid batteries due to slowed chemical reactions.[27] Plugged in overnight, these devices ensure starting voltages of 10–12 volts versus sub-9 volts in unheated conditions, preventing no-start scenarios and extending battery life by avoiding deep discharges from prolonged cranking.[28] They are especially vital for high-compression engines requiring 300–600 CCA for reliable ignition.[29] Intake air and fuel preheating systems, including electric grid heaters or fuel line warmers, raise inlet temperatures by 20–50°C (36–90°F) to enhance fuel atomization and combustion stability, reducing unburned hydrocarbons by over 50% during the first 100 seconds of operation.[30] In diesel fuels prone to wax crystallization below -10°C (14°F), anti-gel additives or heated filters prevent clogging, ensuring consistent delivery.[31] Engine management strategies, such as negative valve overlap via variable timing, trap exhaust residuals for internal reheating, further minimizing emissions without external hardware.[32] Operational practices complement hardware: Idling for 20–30 seconds post-start allows oil circulation before load application, reducing initial wear by ensuring hydrodynamic lubrication films form across surfaces.[33] Coolant flow regulation in modern engines directs heat to critical areas like the cylinder head, accelerating catalyst light-off to 300°C (572°F) faster and curbing transient emissions.[34]Computing and software
Cold boot processes
A cold boot process refers to the complete initialization of a computer system from a powered-off state, triggered by activating the power supply after full shutdown, which ensures all hardware components undergo thorough reset and verification. This method, also termed hard booting, differs from warm booting, where the system restarts via software command without power interruption, thereby bypassing extensive hardware reinitialization and resulting in shorter startup times.[35][36][37] Upon pressing the power button, the power supply unit delivers voltage to the motherboard, prompting the CPU to fetch and execute initial instructions from firmware stored in non-volatile ROM, such as BIOS or UEFI. The firmware conducts the Power-On Self-Test (POST), a diagnostic routine that sequentially checks core hardware—including CPU registers, RAM integrity via memory tests, chipset functionality, and basic I/O devices like keyboard and video output—halting with error signals (e.g., beep codes or LED indicators) if faults are detected. Successful POST clears the CMOS warm boot flag, confirming a full cold start, and proceeds to locate bootable media based on the predefined boot order in CMOS setup.[38][39] The firmware then reads the boot sector from the selected device, typically loading the master boot record (MBR) or GUID Partition Table (GPT) equivalent into memory, which contains the bootloader code. This bootloader, such as GRUB for Linux or Windows Boot Manager, parses the OS configuration, loads the kernel image into RAM, and passes control to it, often after optional stages like loading an initial RAM disk for drivers. The kernel subsequently initializes device drivers, allocates system resources, mounts the root file system, and starts user-space processes, culminating in the desktop environment or shell prompt. In UEFI systems, this sequence incorporates Secure Boot validation to ensure only trusted loaders execute, enhancing security over legacy BIOS modes.[38][40] Cold boot processes are utilized in maintenance scenarios to mitigate transient hardware states, such as clearing volatile memory contents or resetting peripheral controllers that warm boots may not address, thereby resolving issues like unstable connectivity or performance degradation without component replacement.[41] However, the full initialization extends boot duration, often by 10-30 seconds compared to warm boots, depending on hardware complexity and firmware optimizations.[37]Cold starts in serverless and cloud computing
In serverless computing platforms such as AWS Lambda, Azure Functions, and Google Cloud Functions, a cold start refers to the initial latency incurred when invoking a function without an existing warm execution environment, necessitating the provisioning of a new instance including container initialization, runtime setup, and code loading.[42][43] This delay arises from the architecture's emphasis on cost efficiency, where execution environments are terminated during idle periods to avoid charging for unused resources, contrasting with traditional always-on virtual machines.[42] The cold start process typically unfolds in distinct phases: first, the platform allocates and downloads a container image or execution sandbox, which can take tens to hundreds of milliseconds depending on image size and network conditions; second, the runtime environment (e.g., Node.js or Python interpreter) is initialized; third, any extensions or dependencies are loaded; and finally, the function code executes its initialization logic before handling the request.[42] For instance, in AWS Lambda, cold start durations for Node.js functions averaged around 100-500 milliseconds in 2023 benchmarks, though heavier languages like Java or .NET can exceed 1-2 seconds due to longer package loading times.[44] In Google Cloud Functions, HTTP-triggered functions in lightweight languages often exhibit sub-100 millisecond cold starts, outperforming AWS in some comparative tests, while Azure Functions on consumption plans have reported extremes up to 30 seconds in rare scaling scenarios, attributed to sandbox provisioning delays.[45][46] Factors influencing cold start severity include function memory allocation, code package size, runtime choice, and invocation patterns; larger deployments amplify download times, while sporadic traffic exacerbates the issue by increasing the likelihood of environment termination.[42] Empirical studies confirm that cold starts constitute 10-50% of total latency in low-traffic serverless applications, with variability across providers stemming from differences in container reuse strategies—AWS prioritizes aggressive warm-keeping for high-throughput workloads, whereas Azure's sandbox model introduces additional overhead in bursty scenarios.[45][47] In broader cloud computing contexts, similar dynamics appear in containerized services like AWS Fargate or Kubernetes pods without pre-warmed pools, though serverless abstracts these further by fully managing scaling.[48] Mitigation in production environments often involves platform-specific features like AWS Provisioned Concurrency, which pre-initializes instances to cap cold starts at near-zero for predictable loads, or scheduled "ping" invocations to maintain warmth, reducing average latency by up to 90% in tested workflows.[42] Research into advanced techniques, such as function fusion for chaining executions or reinforcement learning for predictive scaling, demonstrates potential reductions in cold start frequency by 30-70% without dedicated concurrency, though these require workload profiling to avoid over-provisioning costs.[49][50] Despite optimizations, cold starts remain a fundamental trade-off in serverless paradigms, prioritizing elasticity over consistent low-latency guarantees suitable for batch or event-driven tasks rather than real-time interactive systems.[47]Performance implications
Cold starts in serverless computing introduce significant latency overhead during function invocation, as the platform must provision a new execution environment, load code and dependencies, and initialize the runtime before processing the request. This overhead typically ranges from hundreds of milliseconds to several seconds, varying by provider, programming language, and configuration factors such as package size and VPC usage. For example, empirical measurements show average cold start latencies under 1 second for AWS Lambda across supported languages, 0.5–2 seconds for Google Cloud Functions, and up to 5 seconds for Azure Functions.[45] The primary performance impact is increased tail latency, particularly affecting p99 metrics in latency-sensitive workloads, where even infrequent cold starts—occurring in less than 1% of requests in AWS Lambda—can degrade overall response times and user experience. This variability stems from resource contention during scaling and inactivity periods, leading to inconsistent throughput; for instance, studies report throughput drops from 470 to 430 requests per second under load due to cold start-induced delays. In real-time or interactive applications, such as APIs or event-driven systems, this can result in perceived slowdowns, reduced scalability, and challenges in meeting service-level agreements.[42][51] Beyond latency, cold starts elevate resource consumption during initialization without productive computation, contributing to higher operational costs as platforms bill for the full execution duration, including overhead. They also limit the suitability of serverless architectures for bursty or low-traffic workloads requiring predictable performance, potentially necessitating hybrid approaches with warm pools or provisioned capacity to mask the effects, though these add complexity and baseline expenses. Research highlights additional downstream issues, including amplified delays in concurrent request scenarios and security risks from extended container reuse post-initialization.[51][52]Recommender systems and machine learning
The cold start problem
The cold start problem refers to the challenge in recommender systems where insufficient historical interaction data hinders the generation of accurate personalized recommendations, particularly for new users, new items, or entirely new systems.[53] This issue arises primarily in collaborative filtering approaches, which rely on user-item interaction matrices to infer preferences through patterns like similarity between users or items; when a user or item enters with zero or minimal ratings, the matrix becomes too sparse to yield reliable predictions.[54] For instance, in matrix factorization models, the absence of data points for cold entities prevents effective latent factor learning, resulting in fallback to generic popularity-based suggestions that fail to capture individual tastes.[55] The problem manifests in three principal forms: user cold start, where newcomers lack prior ratings and thus cannot be profiled against existing users; item cold start, affecting newly introduced products or content with no consumption history; and system cold start, occurring in nascent platforms devoid of any interaction data.[56] Empirical studies quantify its severity, showing recommendation accuracy drops of up to 50% or more for cold users compared to warm ones in datasets like MovieLens, where new users represent 20-40% of interactions in real-world deployments.[57] Causal factors include the data dependency of algorithms—content-based methods mitigate user cold starts via item features but struggle with item cold starts lacking metadata, while hybrids inherit partial vulnerabilities.[58] Consequences extend beyond immediate inaccuracy, fostering user dissatisfaction and higher churn rates; for example, platforms like Netflix or Amazon report that unresolved cold starts contribute to 10-20% early abandonment in streaming or e-commerce contexts. This sparsity exacerbates in dynamic environments with high turnover, such as social media or app stores, where millions of new entities daily overwhelm traditional batch training, underscoring the need for proactive data acquisition without compromising privacy or user experience. Overall, the cold start undermines the core value proposition of personalization, limiting network effects and scalability in data-driven economies.[59]Types of cold starts
In recommender systems, the cold start problem is categorized into three primary types based on the entity affected by the absence of historical interaction data: user, item, and system cold starts. These distinctions arise because collaborative filtering algorithms, which rely on patterns in user-item interactions, fail when data sparsity prevents reliable inference of preferences or similarities.[60][61] User cold start occurs when a new user enters the system without prior interactions, ratings, or profile data, rendering personalized recommendations impossible through methods dependent on historical behavior. This issue affects approximately 40-60% of initial recommendations in platforms like e-commerce sites, where the lack of user-specific data leads to reliance on non-personalized baselines such as popularity-based suggestions. Empirical studies show that user cold start can reduce recommendation accuracy by up to 30% in matrix factorization models until sufficient interactions accumulate, typically requiring 5-10 user actions for stabilization.[62][56][60] Item cold start emerges upon adding new items—such as products, content, or media—to the catalog, which lack user feedback, ratings, or engagement metrics. Without interaction data, similarity computations between the new item and existing ones become unreliable, often resulting in the item being overlooked in rankings and receiving fewer impressions. For instance, in content platforms, new articles or videos may garner 50-70% fewer views initially compared to established items, as confirmed by analyses of real-world datasets like MovieLens, where cold items exhibit normalized discounted cumulative gain (NDCG) scores dropping below 0.2 until 20-50 interactions occur. This type is particularly acute in dynamic environments like news feeds or marketplaces, where inventory turnover is high.[58][63][60] System cold start, also termed community or marketplace cold start, describes the scenario in nascent recommender systems or newly launched platforms with minimal overall user-item interactions, creating a sparse global matrix from inception. This foundational sparsity hampers all recommendation pipelines, as there are insufficient network effects or collective data to seed even aggregate models like item popularity. Historical examples include early-stage two-sided marketplaces, where adoption stalls without initial matches; quantitative evaluations indicate that systems may require 1,000-10,000 interactions to achieve baseline performance, with bootstrapping delays extending launch viability by weeks to months. Unlike user or item variants, system cold start demands external data importation or non-collaborative seeding to initiate feedback loops.[64][61]Strategies for resolution
Strategies for addressing the cold start problem in recommender systems primarily target new users, new items, or both, by leveraging auxiliary data, initial interactions, or advanced modeling to bootstrap recommendations despite limited historical data. For user cold start, where newcomers lack interaction history, common techniques include soliciting explicit feedback during onboarding, such as requesting ratings or selections from a predefined set of items to infer initial preferences.[65] [53] This approach, evaluated on datasets like MovieLens, enables quick profile building with minimal user effort, though it risks fatigue if overdone.[65] Alternatively, recommending popular items favored by established users serves as a non-personalized fallback, improving engagement until sufficient data accumulates, as demonstrated in systems handling sparse user matrices.[65] Active learning methods refine user profiles iteratively by selecting items for rating based on uncertainty reduction or decision trees, reducing the number of queries needed for accurate modeling.[66] Demographic data collection, such as age or location, further aids inference by clustering similar users, though effectiveness depends on correlation strength with preferences.[53] For item cold start, content-based filtering exploits metadata like genres or descriptions to compute similarities with rated items, mitigating sparsity without user interactions.[53] [66] Hybrid systems integrate collaborative filtering with content or demographic signals, enhancing robustness; for instance, combining neural networks adapted for sparse data with classification algorithms like naive Bayes yields better predictions for novel items.[66] Advanced techniques, such as mining discriminant frequent patterns from warm user clusters, update cold user matrices by associating high-frequency itemsets across groups, achieving precisions up to 0.903 on benchmarks like MovieLens 100K.[63] Meta-learning frameworks, which learn initialization parameters from related tasks, have shown promise in recent evaluations for rapid adaptation to cold scenarios, though they require diverse meta-training data.| Strategy Category | Key Techniques | Applicability | Example Performance |
|---|---|---|---|
| User Onboarding | Initial ratings/selections, questions/tags | New users with no history | Reduces queries via active learning [2017 survey][66] |
| Non-Personalized | Popularity-based recommendations | Both user/item cold start | Engagement boost in sparse systems [2024 mapping][65] |
| Content/Demographic | Feature similarity, metadata clustering | Item cold start primary | Handles new items via descriptions [2022 review][53] |
| Pattern Mining | Discriminant itemsets from clusters | In/out-of-matrix cold users | Precision 0.903 on MovieLens [63] |
| Meta-Learning | Task adaptation from priors | General cold start | Improved few-shot learning [2025 review] |