FI Systems, CorpEfficiency is Everything

ABOUT FI SYSTEMS

Research at FI Systems.

Illustration: an electronics research bench
Electronics, batteries and trusted access.

We explore ways to make technology more useful, from everyday electronics to future batteries and trusted access.

FI Efficiency studies ways to help devices run longer and weigh less. Our graphene battery research explores charge retention and useful performance. QNUK investigates one-use quantum credentials.

Explore FI Efficiency →

Explore graphene battery research →

Explore QNUK →

Explore the FI Efficiency research

DEVELOPMENT PORTFOLIO

Tools for demanding electronics.

FI is developing tools in eight product areas for defense, industrial and corporate teams, and commercial manufacturers. Each area targets less weight, longer battery life, lower power use or less calculation while preserving the product’s required functions.

Illustration: Sensor reading plans illustration
Reading plans for instruments and connected sensors.
In development

01 / FI SYSTEMS, CORP

Sensor reading plans

Plan how many readings each sensor takes.

Fewer readings at the same quality, or a steadier answer.

Who it serves

Teams designing measuring instruments and sensors.

Intended use

One model choice takes fewer readings at the required quality, with added planning energy counted. The other keeps the original readings for better precision.

Where this could help
  • Defense: Portable sensors could run longer or carry a lighter battery at the same sensing quality.
  • Industry and business: Inspection and monitoring teams can compare fewer readings at fixed quality with their battery use and the cost of taking each reading.
  • Commercial products: For mobile and connected sensors, the choice under study is lower task energy or better precision with the original number of readings.
Results and remaining checks

3.36–29.94% longer battery life per charge, or a lighter sensing design.

The starting plan assigns equal reading counts. Exact greedy allocation and exhaustive search match FI’s allocations. In the best selected case, 48 readings become 23 while meeting the old quality limit. Under the stated energy model, a job falls from 1 J → 0.7696 J, equivalent to 1 W → 0.7696 W at one job per second.

MeasureResultComparison
Battery life per charge: keep the same battery3.36–29.94% longer10 hours → 10.34–12.99 hours at the same sensing rate and usable battery. Alternative to downsizing the battery.
Power / energy3.25–23.04% lowerEight qualifying cases versus equal reading counts; same useful work rate.
Heat3.25–23.04% less per jobNormalized heat projection: equal exports and stored energy; compared energy becomes heat.
Passive cooling: surface area3.25–23.04% smaller in ideal sizingAt fixed heat transfer per effective area and temperature difference, required passive surface area scales with the thermal-design heat load.
Cooling assembly size / weight17.28% lower in the worked design200 g → 165.44 g and 200 cm³ → 165.44 cm³. The 50 g / 50 cm³ fixed portions stay; the 150 g / 150 cm³ variable portions scale with the best-case 23.04% load reduction.
Device: battery + cooling resize9.22% lighter; 10.37% smaller in the worked design1,000 g → 907.83 g; 1,000 cm³ → 896.31 cm³. Fixed original battery life per charge, unchanged densities and packing, no added parts. These are component-budget calculations.
Battery portion: size / weight3.25–23.04% lower in ideal scalingAlternative to longer battery life per charge: fixed original battery life per charge, unchanged battery energy density and packing, no added parts. This applies to the energy-bearing battery portion, not the whole device.
Reading work8.33–52.08% fewer readingsNine saved cases that improved the minimum reading count while preserving the old quality limit.
Precision at the old budget26.30% lower mean varianceAlternative choice: retain the original readings; mean across all 12 selected cases.

Readings use 50% of the original job energy, and added processing costs 3%. The calculation assumes omitted readings are physically skipped while energy per reading and the useful reporting rate stay fixed. A design can spend the saving on a smaller battery or longer battery life per charge.

PASSIVE-COOLING SIZE CALCULATION

A smaller cooling assembly.

These figures show how the saved sensing result changes passive cooling at steady state.

MeasureBeforeAfterDifference
Design heat load10 W7.7 W23.04% lower
Effective passive-cooling area500 cm²384.79 cm²23.04% smaller
Cooling assembly mass200 g165.44 g17.28% lighter
Cooling assembly volume200 cm³165.44 cm³17.28% smaller
Device mass: cooling resize only1,000 g965.44 g3.46% lighter
Device volume: cooling resize only1,000 cm³965.44 cm³3.46% smaller
Device mass: battery + cooling resize1,000 g907.83 g9.22% lighter
Device volume: battery + cooling resize1,000 cm³896.31 cm³10.37% smaller

Illustrative 1 kg, 1 L design: 200 g and 200 cm³ of passive cooling, with 50 g and 50 cm³ fixed; a 250 g, 300 cm³ scalable, energy-bearing battery portion. The combined case resizes the battery for the original battery life per charge.

Calculation and sources

Convection heat balance: Q = h × A × ΔT. At fixed heat transfer per area and surface temperature, area scales with design heat load. With assumed h = 10 W/m²K and ΔT = 20 K, 10 W needs 500 cm²; 7.6958 W needs 384.79 cm².

The 150 g and 150 cm³ variable cooling portions scale with effective area; the 50 g and 50 cm³ fixed mounting/base portions remain. Thus cooling mass = 50 + 150 × (1 − 553/2400) = 165.4375 g. Cooling-only device saving = 34.5625 g and 34.5625 cm³.

For fixed original battery life per charge and unchanged battery energy density: combined mass saving = (150 + 250) × 553/2400 = 92.1667 g; combined volume saving = (150 + 300) × 553/2400 = 103.6875 cm³. The added-part budget is zero; processing energy is already included in the net reduction.

The example assumes the thermal-design load falls with average energy, stable heat-transfer effectiveness, equivalent materials and packing, unchanged ambient and temperature limits, and the same rate of required work. The modeled resize also assumes compatible peak heat, contact resistance, airflow clearances and structural support.

Radiative area has the same proportional relationship at fixed emissivity, surface/sink temperatures and external heating. Whole-device volume uses the assumed removable component volumes.

Review scope: Three equal-noise cases save no readings and use 3% more energy after added processing.

Technical research record
Illustration: Shared errors in measurements illustration
Related measurements and the errors they share.
In development

02 / FI SYSTEMS, CORP

Shared errors in measurements

Model errors that several measurements have in common.

See how related errors shape the answer.

Who it serves

Teams combining sensor readings, calibrating instruments or building numerical software.

Intended use

Related measurements can share the same error. Modeling that link helps assess the quality of the combined answer.

Where this could help
  • Defense: Equipment with several sensors: study reading counts with shared errors included.
  • Industry and business: Inspection teams and teams combining sensor readings can assess measurement quality with linked errors included.
  • Commercial products: Calibration software can compare sensing quality when a smaller hardware design is being considered.
Results and remaining checks

Retained C05 summary ratios imply 41.40–97.44% less calculation time.

T032’s retained C05 summary compares selected 256-input cases with SymPy. Converting its ratios into time gives 58.5995–2.5599 seconds for a normalized 100-second reference. At an assumed constant 1 W, the corresponding energy calculation is 100 J → 58.5995–2.5599 J. The shared-input uncertainty model remains in the calculation.

MeasureResultComparison
Compute time41.40–97.44% lower from summary ratiosCalculated from T032’s retained SymPy/C05 summary ratios of 1.7065–39.0642 for selected 256-input cases, described as within-case medians of three invocations.
Task energy / heat41.40–97.44% lower from summary ratios in a constant-power projectionSame compared calculation and unchanged active power; heat assumes the compared energy becomes heat. No measured energy reduction.
Answer checks4,096-observation match reported in the retained T031 summaryThe retained T031 summary reports an exact-answer and separate certificate check in its restricted synthetic model.

Both sides perform the same exact inference task with the same result requirements. The selected historical comparisons use a restricted shared-input model. Converting time to energy separately assumes constant active power.

Review scope: Original T032 invocation witness unrecovered; the ratios come from the retained summary.

Technical research record
Illustration: Amplifier models illustration
Amplifier behavior in radio and instrumentation.
In development

03 / FI SYSTEMS, CORP

Amplifier models

Predict amplifier output from current and earlier inputs.

Plan signal correction with processing costs in view.

Who it serves

Radio, telecom and instrument developers.

Intended use

Compare amplifier predictions as operating conditions change, before choosing signal correction and the processing needed to run it.

Where this could help
  • Defense: Scanning, spoofing and jamming studies compare correction costs at the same signal effect and duty cycle.
  • Industry and business: Teams designing radio-frequency instruments can compare signal-correction methods and processing costs.
  • Commercial products: Wireless and cellsite designers can compare correction costs while keeping the same delivered service.
Results and remaining checks

99.986–99.9999% less prediction error in the synthetic model.

The generated amplifier model changes operating state over time. Including its recent state history reduced prediction error during slow, fast and intermediate changes. The simplest model used four features; the history model used twelve. The test plant was synthetic and known.

MeasureResultComparison
Prediction quality99.986–99.9999% lower mean squared errorThree generated test regimes; 3,000 test samples per regime.
Model complexity4 → 12 featuresThree times as many model features; execution time and energy unmeasured.

6,000 training samples; three separate generated test regimes. The task predicts the synthetic plant’s forward output.

Review scope: The history model uses more features. Its net processing and energy costs remain unmeasured.

Technical research record
Illustration: Controller-code checks illustration
Controller code checked against its design reference.
In development

04 / FI SYSTEMS, CORP

Controller-code checks

Check code generated for a device.

Review controller changes before they reach hardware.

Who it serves

Teams writing software for devices and automated equipment.

Intended use

Compare generated control logic with the design reference before it goes into a device.

Where this could help
  • Defense: Field-equipment studies can check control behavior while exploring smaller hardware or lower power.
  • Industry and business: Factory-control and robotics developers can check proposed changes against required output and deadlines.
  • Commercial products: Connected-device and energy-controller developers can compare control logic before putting it into hardware.
Results and remaining checks

Check controller behavior before hardware integration.

FI’s development workflow compares the generated logic with its design reference before device integration.

MeasureResultComparison
Correctness45 matching sequences reported in the retained controller summaryCorrected generated controller versus its original reference logic.

The retained FI controller summary reports matching reference outputs for 45 specified input sequences.

Review scope: The result covers the 45 specified input sequences in the retained summary; the original execution witness is unrecovered.

Technical research record
Illustration: Power and heat analysis illustration
Energy flows, heat and complete power budgets.
In development

05 / FI SYSTEMS, CORP

Power and heat analysis

Trace power from the source to the useful output.

Find where energy becomes unwanted heat.

Who it serves

Teams designing electronics, power converters and equipment.

Intended use

Count losses in the source, conversion equipment and load, including added demand, at the same useful output.

Where this could help
  • Defense: Teams building portable or vehicle-mounted electronics can compare battery weight, runtime and power drawn at the same output.
  • Industry and business: For motors, pumps and controls, the comparison traces energy losses at the same force, flow or production rate.
  • Commercial products: Phone, solar-control and telecom studies compare battery life, electricity use and heat across the full power budget.
Results and remaining checks

Source-to-load energy checks.

The model tracks energy at the load and battery, including conversion losses and losses from current variation. In the equal-force calculation, reducing ripple alone leaves copper loss unchanged when the required force is preserved.

MeasureResultComparison
Energy boundaryLoad-side saving can reverse at the batteryOne retained model shows 4.02 mJ less at the load while battery use increases by 6 mJ.

The electrical and actuator models keep the same useful output. The budget includes the full power-source energy cost.

Review scope: The rejected cooler schedule used more energy and took longer to complete.

Technical research record
Illustration: Signal recording plans illustration
Signal samples, timestamps and operating context.
In development

06 / FI SYSTEMS, CORP

Signal recording plans

Plan what a signal recording needs to keep.

Know what the recording needs to preserve.

Who it serves

Receiver, instrument and test teams.

Intended use

Define the required samples, timestamps and operating context before considering smaller recordings.

Where this could help
  • Defense: Scanning and authorized spoofing/jamming studies can compare recorder storage while keeping the required signal evidence.
  • Industry and business: Instrumentation and interference-log designs can size storage for the complete required record.
  • Commercial products: Wireless and telecom recorder studies compare more recording capacity or smaller storage, keeping the required samples and context.
Results and remaining checks

Smaller-recording candidate failed retention quality.

The smaller recording was compared under the same evidence-retention and downstream analysis requirements.

MeasureResultComparison
Record qualitySmaller-recording candidate rejectedThe retained RF1 summary reports lost data required by the specified recording-quality requirement; the original execution witness is unrecovered.

Same evidence-retention requirements and downstream analysis quality.

Review scope: The retained RF1 summary reports loss of required data. The original execution witness is unrecovered.

Technical research record
Illustration: Offline engineering workspace illustration
Local models, calculations and review records.
In development

07 / FI SYSTEMS, CORP

Offline engineering workspace

Keep models and results available without a live network.

Reuse unchanged calculations. Review them offline.

Who it serves

Field, factory and product engineers working offline.

Intended use

Compare designs locally, reuse unchanged calculations and keep the assumptions and checks with the results.

Where this could help
  • Defense: Scanning, spoofing and jamming studies can compare kit weight, runtime and vehicle power at the same requirements.
  • Industry and business: Factory and corporate engineers can reuse calculations and simplify history handling while keeping the checks.
  • Commercial products: Product engineers can reuse unchanged results and weigh memory cost against task time and energy.
Results and remaining checks

Keep engineering review moving when connectivity is limited.

The proposed workspace keeps design models, calculation inputs and review results together locally. Engineers could compare alternatives, retain assumptions and revisit results while disconnected from a live network.

MeasureResultComparison
Exact report-helper math23.08% fewer expression evaluations3,194,940 → 2,457,660 in the source-derived count, including 60 final divisions. Conventional implementations get the same cache; 0% FI-exclusive saving against that baseline.
Named repeated work100% removed in two categoriesAt 4,096 records, those categories accounted for 16,777,216 visited positions in the earlier calculation workflow.

The examples cover repeated history and exact report arithmetic. At 4,096 records, two removed categories account for 16,777,216 visited positions. Separately, report expression evaluations fall from 3,194,940 to 2,457,660 (737,280 fewer; 23.08%). Expression evaluations and history traversals are separate work counts.

Review scope: Caching keeps intermediate values in memory. Conventional software with the same cache ties the arithmetic count.

Technical research record
Illustration: Receiver logs and fault review illustration
Signal records alongside receiver operating logs.
In development

08 / FI SYSTEMS, CORP

Receiver logs and fault review

Connect signal events with equipment logs.

Trace the context of a suspected signal fault.

Who it serves

Communications, instrument and equipment-review teams.

Intended use

Bring recorded signal intervals together with receiver state so reviewers can investigate a suspected fault or interference event.

Where this could help
  • Defense: Portable receiver and scanning-system studies keep fault context when studying changes to hardware or power use.
  • Industry and business: Instrumentation teams can connect faults with operating logs when reviewing equipment changes.
  • Commercial products: Wireless and telecom studies can keep traceable signal and equipment records while considering smaller hardware.
Results and remaining checks

Receiver logs and signal records are being linked for review.

The planned workflow links suspect signal intervals to the relevant receiver logs. Its comparison is an ordinary review process covering the same events, retained evidence and false-alarm burden.

Same receiver inputs, event labels, required evidence and review task.

Review scope: Detection accuracy, review-time savings and device efficiency remain unestablished.

Technical research record

OUR DESIGN APPROACH

Keep the result. Reduce the work.

FI uses arithmetic libraries for number operations, then adds choices tailored to the job. It recognizes exact input patterns, takes a shorter calculation path where available, checks the required result and keeps a general fallback.

Specialized calculations

In H22’s 2,500 cases, choosing shorter exact paths cut the sum of case-median times by 3.54% against the prior FI hybrid: a normalized 100 → 96.46 seconds. The matching-input group used 31.88% less calculation time. Four of ten groups improved and six slowed. A separate 15 October independent AI accounting review found FI slower than capable ordinary gmpy2 in all 26 preserved first-call and warm timing groups.

Shared history

The revised offline design stores checked, unchanged history once, so calculations can access it without repeatedly copying and rereading the growing record. At 4,096 records, the two removed work categories accounted for 16,777,216 visited positions. Full result checks remain. A separate report design caches exact intermediate values, cutting expression evaluations from 3,194,940 to 2,457,660 (23.08%). Both figures are work counts derived from source; conventional software can use the same caching.

Reading allocation

FI assigns readings according to their contribution to the required answer. Across 12 selected model cases, meeting the old equal-plan quality limit needed 23.8% fewer readings on average. If their energy saving exceeds added processing, skipping readings could extend battery life per charge or allow a smaller battery. Conventional exact planners reached the same allocations. FI’s work brings planning, result checking and device-budget analysis together.

Available arithmetic library: gmpy2

The 35-device review follows these changes into application budgets for size, weight, power, heat, computing, speed, battery life per charge and communication range. The linked records give the inputs and cost calculations.

READING-PLAN RESULTS

Reading plans.

Our reading-plan study compared two choices: 23.8% fewer minimum readings at the old equal-plan quality limit, or 26.3% lower modeled variance at the same reading budget, on average across 12 selected cases. Lower variance means a steadier combined answer.

Exact allocation methods matched FI in these cases. The linked energy model assigns 50% of job energy to readings and 3% to added planning.

Calculations and sources
Illustration: a compact sensor assembly, illustrating a possible measurement setting
26.2972% lower modeled variance of the combined estimate, averaged across 12 selected cases.

26.2972% lower modeled variance. A choice in how to use it.

Across 12 selected cases, FI with a shared whole-reading adjustment cut variance by 26.2972% on average versus equal counts at the same budget. Mean standard deviation fell 15.0136%. Exact minimum reading budgets fell 23.7847% on average when matched to the old equal-plan variance limit. The alternatives are better precision at the original reading budget or fewer readings at the old quality limit.

Nine cases improved and three tied at the original budget. Exact greedy allocation and full search matched FI in every case.

Illustration: an electronics test setting, illustrating the need to check results
The retained FI research summary reports arithmetic matches across 1,018 paired rows and 12,200 sum steps. Retained summary-level evidence.

10 report pairs in the retained research summary.

A saved FI research summary reports arithmetic matches across 10 report pairs, 1,018 paired rows and 12,200 running-sum steps. Nine cases were labeled synthetic and one recorded data. That recorded-data label has not been independently verified, and the original report files remain unchecked. This page uses the saved summary.

Retained research summaries and evidence scope
See all 12 model cases and how the average is calculated

Each of four designed variance sets was checked at reading budgets of 12, 24 and 48. The model assumes independent readings within and across channels, fixed variance for each reading, equal cost per reading, and no shared drift or systematic error (a bias in the readings). Variance measures spread; mV² means squared millivolts.

One-reading variances (mV²)Reading countEqual-plan variance (mV²)FI-plan variance (mV²)Reduction
1,1,1123/43/40.00%
1,1,1243/83/80.00%
1,1,1483/163/160.00%
1,4,9127/23/114.29%
1,4,9247/43/214.29%
1,4,9487/83/414.29%
1,9,811291/4113/837.91%
1,9,812491/8113/1637.91%
1,9,814891/16155/4438.06%
1,1,1001251/212/152.94%
1,1,1002451/46/152.94%
1,1,1004851/83/152.94%

For each row: reduction = (equal-plan variance − FI-plan variance) ÷ equal-plan variance × 100. The exact sum of the 12 row percentages is 5,370,000/17,017. Divide by 12 to get 447,500/17,017%, or 26.2972% when rounded to four decimal places. Converting each case before averaging gives a 15.0136% mean standard deviation reduction. Exact lower-budget planning gives 23.7847% mean reduction in minimum readings to meet the old equal-plan variance. Each case has equal weight. Results cover the selected mathematical cases.

A separate, preselected 13-reading case used variances 1, 5 and 13 mV². The best balanced plan, with counts 4, 4 and 5, gives 41/10 mV². FI plus the shared adjustment, with counts 2, 4 and 7, gives 101/28 mV²: 12.0209% less modeled variance. Both exact methods tie it. This case is excluded from the 12-case mean.

Every allowed positive whole-reading split was checked: 5,556 across the 12 exploratory cases and 66 in the separate case. Exact mathematical allocation checks.

A CLOSER LOOK AT THE EVIDENCE

Same reading count.
A different plan.

Where each reading goes can change the result. Explore three saved model cases, including one with no gain.

Different spread, 13 readings

MODELED CALCULATION

How the readings are shared

One square is one reading. Both plans use 13.

Best balanced splitFI + shared adjustment
Channel 1
Channel 2
Channel 3
12.02%less modeled variance
than the best balanced split

Variance measures how much the answer can vary; a lower value means a steadier modeled answer.

Exact comparison methods reach the same result. FI adds 0% extra variance reduction over them.

The readings are independent in this model, with a fixed spread and equal cost per reading. Shared drift and systematic bias are excluded.

View the calculation

For each channel, divide its one-reading variance by its reading count, then add the three values.

Best balanced split: 1/4 + 5/4 + 13/5 = 41/10 mV²

FI + shared adjustment: 1/2 + 5/4 + 13/7 = 101/28 mV²

Reduction = (balanced-plan variance − FI-plan variance) ÷ balanced-plan variance × 100 = 12.02%, rounded.

A balanced split keeps channel counts within one reading of each other. For 13 readings, the baseline is the best of the three balanced splits.

The shared adjustment assigns whole readings. Exact greedy allocation and a search through every allowed split reach the same result in these selected examples.

How we assess the evidence

WHAT LESS WASTE CAN MAKE POSSIBLE

10 hours → 12.99 hours.
Or 92.17 g less mass.

The best selected sensor result gives two choices in the 1 kg model: keep the battery to run longer, or make the battery and passive cooling smaller. Both include the added 3% processing energy.

Illustration: compact processor with cooling fins
Less energy lost as heat can ease the burden on cooling.

Give the cooling system less to do.

Lower heat generation may allow longer continuous use in heat-limited sensors and controllers.

Illustration: precision actuator and electronic control equipment
Saved energy could support extra sensor checks or control updates.

Get more from the same energy budget.

Energy saved by one task could instead power another sensor, more signal processing or extra control updates.

Go longer between charges.

The eight selected energy cases give 3.36–29.94% longer battery life per charge at the same useful sensing rate and usable battery. In the equal-noise loss cases, time between charges falls 2.91%.

Make room for a lighter design.

Keeping the original time between charges, the model gives 1.30–9.22% less weight and 1.46–10.37% less volume. Those figures use the stated battery and cooling shares, with the smaller parts assumed to fit and be removable.

Illustration: receiving antennas for passive radio research
Keep signal records and receiver logs together for review.

GOVERNMENT & DEFENSE

Passive sensing.

We develop reading plans, recording methods and linked receiver logs for passive sensing. Interference research covers system requirements, test methods and evidence.

Passive-sensing research in development.

Discuss an application