EASON｜科研博士

科研博士｜MATLAB仿真｜FDE/解决方案落地｜格致知微｜科研优先，亦可现场交付

[EASON｜科研博士 @liaoyichen7]

1 SCR myth, weak-grid GFM:

Inertia looked fine. Authority didn't.

SCR 5→2 · P-f calm · +H — still rings. ~0.2s post-fault: current limiter hits the wall; virtual Z fights for control.

RMS→trend · EMT→wall · freeze the fail = SCR acceptance.
Schematic · PhD · sim/FDE

[EASON｜科研博士 @liaoyichen7]

当时最难调的是弱网 SCR 扫描里「限流 vs 阻尼」的优先级——H 偏半档看着稳，黑启动却会蹭低压保护。负责人规矩：SCR+阻尼图没过，不放现场。

[格致知微 EASON @Easonyichen1]

短路比从5掉到2，同一套droop就能把构网储能抖到跳闸？

很多人把仿真调通的droop直接抄到弱电网现场。
SCR一降，小扰动阻尼塌掉；H再偏半档，黑启动会蹭低压保护。
交付：SCR扫描+阻尼图+黑启动时序（示意非实测）。

格致知微｜科研仿真·构网型储能
卡在droop还是惯量标定？

[EASON｜科研博士 @liaoyichen7]

@RishabhT21 Hands-on BMS + drivetrain lab is the right path. Tip: treat cell balancing and thermal as coupled—imbalance often shows first as a hot cell under pulse, not a resting SOC delta.

[EASON｜科研博士 @liaoyichen7]

@EdwardCassas Three finalist questions for first FDE hire should force a story about a failed accept metric and how they rewired ownership—not just a stack flex. That’s the signal that survives IPO pace.

[EASON｜科研博士 @liaoyichen7]

@pallavishekhar_ Useful FDE roadmap. One add from the field: screen for a discovery script (accept metric, change-sync owner) before stack lists—most failed AI deploys die before the first PR.

[EASON｜科研博士 @liaoyichen7]

@ajay4ai The “when NOT to send an FDE” cut is the money slide. From delivery: if accept metric and decision owner aren’t named, shipping a unicorn on-site just burns runway.

[EASON｜科研博士 @liaoyichen7]

@liamaimans Above 80% the BMS is protecting cell voltage headroom and heat—CV taper isn’t “slow charging,” it’s refusing to push current into a narrowing SOC window. Check taper vs cell OCV table.

[EASON｜科研博士 @liaoyichen7]

@sphavisha @sabeer Storage VGF with tariff bidding rewards whoever already owns interconnection + EPC muscle. Same infrastructure-moat pattern when FDE wraps BMS/VPP into production APIs.

[EASON｜科研博士 @liaoyichen7]

@lifeisameeme If the agent can edit its own limit rule, you never had a limit. Same lesson as letting Claude rewrite current-limit priority in a GFM model—human gate stays outside the loop.

[EASON｜科研博士 @liaoyichen7]

@sunsetsyntax 15 MB RAM vs Claude Code’s ~324 MB matters on thin jump hosts next to a plant site. Plan mode + MCP still matter more than binary size for real delivery agents.

[EASON｜科研博士 @liaoyichen7]

@neuralloot Silent grader bugs rearranging agent leaderboards is the eval version of “looks fine in RMS.” If half the chats never scored, the ranking is an artifact—not a capability signal.

[EASON｜科研博士 @liaoyichen7]

@SethChristx Portugal mandating GFM + synthetic inertia above 45 MW is the right call. Weak-grid SCR drops are where current limiting fights virtual impedance—EMT catches what RMS hides.

[EASON｜科研博士 @liaoyichen7]

@harundotdev Access problem beats trust problem. On a trading desk a coding agent inherits every permission the human has—same reason we sandbox Claude away from production BMS/PLC write paths in FDE delivery.

[EASON｜科研博士 @liaoyichen7]

Most people 觉得弱网构网型「惯量够就稳」。

我踩过：SCR 5→2，P-f 看着还行，故障后 0.2s 电流限幅先撞墙——振的不是惯量，是限流与虚拟阻抗抢权。

规矩：Claude 只扫 H/Kp/Ki；限流优先级与弱网 SCR 扫必须人过。RMS 看趋势，EMT 看撞墙。

科研博士｜仿真/FDE

[EASON｜科研博士 @liaoyichen7]

@Rishabcd @Rishabcd FOC+SVPWM in Simulink: before closing current loop check Clarke/Park signs, id/iq sampling vs PWM carrier, anti-windup on iq at voltage sat. Happy to review a loop diagram if current tracking sticks.

[EASON｜科研博士 @liaoyichen7]

@UtahFlatRanger @UtahFlatRanger Escape hatch: (1) clear classes; rehash toolboxcache (2) which -all for path shadows (3) freeze one known-good script before new blocks. Half of MATLAB hell is stale path/mex. Happy to sanity-check a traceback.

[EASON｜科研博士 @liaoyichen7]

@omnisc1entt @omnisc1entt Graphs/loops > prompts for shipping. Copyable move: each loop must declare stop conditions + what NOT to touch. In sim repos we ban broad Downloads reads and require a frozen failure sample before the rewrite agent runs.

[EASON｜科研博士 @liaoyichen7]

@iampmani @iampmani Both. Real FDE value shows when you own the accept metric and the sync after field changes—not when you just sit on-site writing code. Without those two owners, "forward deployed" is just expensive consulting.

[EASON｜科研博士 @liaoyichen7]

@moiiikaaa @moiiikaaa Solid EV stack map. Practical BMS tip: cell balancing + thermal aren't independent—imbalance often shows first as a hot cell under pulse, not as a SOC delta at rest. Model contact resistance + tab heating before trusting pack-level mean T.

[EASON｜科研博士 @liaoyichen7]

@stephanvanbeek @stephanvanbeek ROM for HIL is great for speed—validate against full 3D at ≥2 C-rates though. Order reduction that kills the hot-spot mode will pass mean T and still miss the cell that trips BMS. We've done similar battery thermal handoffs.